Five siphons: the AI infrastructure wealth transfer

Abstract

The official story of the AI boom is a story of investment: hundreds of billions of dollars deployed by a handful of well-capitalized companies in a competitive race to build data centers, fabs, and power generation. The unofficial story, the one that gets harder to ignore the closer you look at the books, is that the capital being deployed is not actually theirs. It is being drawn, by five structured channels, from the rest of the economy. None of these channels are illegal. All of them are working as designed. Taken together they represent the largest cost-externalization exercise in modern American infrastructure history, and the largest transfer of wealth from the public to a small number of private balance sheets since the railroads.

The capex setup

Microsoft, Google, Meta, and Amazon will collectively spend approximately $725 billion in capex in 2026, a 77% increase over 2025's record $410 billion. Goldman Sachs's broader AI-capex aggregate, which includes the additional non-Big-Four AI-infrastructure spend (Oracle, Tesla, the model labs' direct compute outlays, and other hyperscaler-adjacent issuers), projects $7.6 trillion in cumulative AI capex between 2026 and 2031, doubling from ~$765 billion in 2026 to ~$1.6 trillion by 2031 on that broader basis. The 2027 number is expected to clear $1 trillion for the first time in any single industry's history.

Goldman is also blunt about what the math implies on the depreciation side: shortening the modeled economic life of AI chips from five years to three (closer to actual hyperscaler-grade utilization) adds roughly $1 trillion to annual industry depreciation expense, which absorbs most or all of the optimistic-case revenue growth the capex commitment depends on (Goldman Sachs Global Investment Research, "Tracking Trillions"). Separately, Goldman analysts have warned that hyperscalers may realize only about half the profit needed to justify their AI capex (Fortune, January 2026), against a 2026 consensus revenue estimate of $450 billion for the relevant cluster.

That gap has to close somewhere. The historical playbook is one of three things: revenue catches up, costs shrink, or impairments destroy shareholder value. Which outcome lands stays unresolved; the wealth transfer that funds the bet operates regardless.

The commercial state of play, as best we can tell

The aggregate numbers obscure who is actually making money. On the vendor side, the picks-and-shovels seller is winning enormously; the model-frontier prospectors are still burning capital at a historic clip, with one notable exception. On the enterprise customer side, the realized ROI is now bifurcated rather than uniformly bad, concentrated in a handful of narrow functions, still distant in most others.

Vendor side. NVIDIA's Q1 FY2027 (calendar Q1 2026) net income was $58.3 billion in a single quarter, on $81.6 billion in revenue, with Data Center revenue of $75.2 billion up 92% year-over-year (NVIDIA Q1 FY27 earnings release, May 2026). If that single-quarter run-rate were to hold steady, the implied annualized net income from one firm would be in the ~$230 billion range. That's a four-times extrapolation, not company guidance; the foundation number is the reported $58.3 billion quarter. By contrast, OpenAI is at approximately $2 billion in monthly revenue ($24 billion annualized) but projecting a $14 billion loss in 2026 and a $27 billion cash burn against an even larger projected loss in 2027; the company's own filings tell investors to expect material losses through 2028 before a forecast turn to profitability in 2030. Anthropic crossed $30 billion annualized revenue in April 2026 and roughly $47 billion by late May 2026 (Simon Willison covering Anthropic disclosures), up from $9 billion at year-end 2025, with cash burn forecast to drop to ~1/3 of revenue in 2026 and ~9% by 2027, the only major model-lab on a credible near-term path to positive cash flow.

Enterprise customer side. The 2024 MIT NANDA finding that 95% of organizations were getting zero return on GenAI pilots set the bear case. The 2026 picture, per McKinsey Global AI Survey 2026 and concurrent BCG / Gartner / Slack Workforce Index data, is more nuanced: production deployments with telemetry are showing real gains, but the gains are highly concentrated by function and the broad-transformation projects still fail at high rates (Gartner: >40% of agentic AI projects will be canceled by end of 2027).

The AI commercial thesis is partially validated in narrow production deployments and broadly unvalidated at the enterprise-transformation level, which is the level the capex commitment requires. NVIDIA's profits are real, very large, and concentrated in one company. OpenAI's losses are real, very large, and externalized through equity-fundraising onto investors and through capex flow-through onto the bondholder pension funds (Siphon 4). The commercial state of play does not foreclose either outcome (revenue catches up; or impairments destroy shareholder value), but it does foreclose the assumption that broad commercial success is already happening at the scale required to justify the capex. It is not.

The cloud capex thesis depends on a second condition that is even less discussed than the ROI gap: that frontier AI hardware stays expensive enough and energy-hungry enough to keep useful inference inside cloud datacenters. That condition is being demolished in 2026, and the company demolishing it is the same one selling the cloud the picks and shovels it bought to enforce it: NVIDIA.

The hardware moat NVIDIA is selling against itself

The cloud-AI business model is, structurally, a hardware-rental business. OpenAI and Anthropic charge per-million-tokens for inference because, until recently, the alternative, owning the hardware to run frontier large language models locally, was infeasible for most buyers. The cloud capex thesis assumes that this remains true: that the hardware required to run useful AI stays expensive enough to keep enterprises, downstream businesses, and consumers dependent on per-token cloud billing. That assumption is being demolished in 2026, and the company doing the demolition is NVIDIA itself.

The pivotal move was GTC Taipei on 1 June 2026, co-timed with Computex 2026, where Jensen Huang unveiled the RTX Spark superchip, the N1X SoC co-designed with MediaTek (NVIDIA Newsroom, "NVIDIA and Microsoft Reinvent Windows PCs for the Age of Personal AI"; MediaTek press release on RTX Spark collaboration). The specifications:

• 20-core ARM-based CPU (10 Cortex-X925 performance cores + 10 Cortex-A725 efficiency cores, ~4.1 GHz peak) paired with a Blackwell RTX GPU (6,144 shader cores) in a two-chiplet SoC on TSMC 3nm.
• Up to 128 GB unified LPDDR5X memory.
• 6,144 CUDA cores; approximately 1 petaflop of FP4 AI compute.
• TSMC 3nm process.

The strategic move that matters: RTX Spark ships this fall in Windows laptops and compact desktops from Microsoft Surface, ASUS, Dell, HP, Lenovo, and MSI, NVIDIA's own statement projects over 30 laptops and more than 10 desktops in the initial wave, including named designs like the Dell XPS 16 Creator Edition, HP OmniBook, and the Microsoft Surface Laptop Ultra. NVIDIA's marketing claim for the platform: capable of running models of 120 billion parameters with 1-million-token context locally. This is not a developer-tier product. It is a consumer-tier product, on every major OEM's lineup before end-2026, picking up the Windows-on-ARM market opportunity that Qualcomm's exclusive Microsoft deal failed to capitalize on.

The multi-generation roadmap is the signal that this is not a one-off. NVIDIA outlined three RTX Spark generations at Computex: Fall 2026 (Grace + Blackwell + LPDDR5X), 2028 (Vera CPU + Rubin GPU + LPDDR6), 2030 (Rosa CPU + Feynman GPU). Each iteration extends capability while keeping the price band consumer-accessible.

The 2025 precursor matters for context. NVIDIA announced "Project DIGITS" at CES 2025 and shipped it as DGX Spark in late 2025 at $3,999 to $4,699, the same Grace+Blackwell silicon, targeted at Linux and AI developers rather than Windows consumers. With 128 GB unified memory, DGX Spark fits the full 70B weight set at 8-bit (INT8/FP8) precision (~70 GB) with comfortable headroom for KV-cache and context, or runs the same 70B at Q4 with very long context, or fits mid-60B-class models at BF16. (Full FP16 weights for a 70B model require roughly 140 GB, which exceeds the 128 GB envelope; that workload belongs to the dual-Spark or H100 tier above.) RTX Spark is the consumer-Windows layer on top of the same silicon, gaming and creation stack added, OEM channel mobilized, mass-market price-band targeted.

The workstation flagship completes the lineup: NVIDIA RTX PRO 6000 Blackwell with 96 GB ECC VRAM, shipping from Dell, HP, Lenovo, and Razer. At FP8 precision on the RTX PRO 6000, Llama 3 70B occupies ~70 GB and leaves 26 GB for KV cache, enough for long-context inference at near-cloud quality on a single workstation card.

The competitive landscape filled in around NVIDIA's moves: AMD's Ryzen AI Max+ 395 ("Strix Halo") ships in 128 GB mini-PCs and laptops in the $2,500-$3,000 street price range (AMD-branded reference at ~$3,999) with up to 96 GB allocatable to the iGPU. Independent llama.cpp / Vulkan / RADV benchmarks (Level1Techs forum, llm-tracker.info, kyuz0 grid-view) report Llama 3.1 70B Q4 at the memory-bandwidth ceiling for that platform and the Qwen3-Coder 30B family running 70 to 100 tok/s depending on backend and quantization (Q4_K_M on ROCm 7 at 70 tok/s; Vulkan/RADV at higher). NVIDIA's consumer RTX 5090 ($2,000) ships with 32 GB GDDR7 and runs 70B Q4 with comfortable context. Apple's high-VRAM M-series workstation silicon rounds out the prosumer field. The 2025-to-2026 progression is unambiguous: every major silicon vendor is now putting frontier-LLM-capable hardware in the consumer Windows / mini-PC / laptop tier.

Open-weight 2026 models have closed the quality gap to cloud frontier:

• Llama 4 Scout (Meta, April 2025): 17B active / 109B total MoE, 10M-token context, fits Q4 on a single 80 GB GPU.
• DeepSeek V4 Flash (April 2026): 284B MoE / 13B active per token, 1M context, MIT license.
• Qwen 3.6-35B-A3B (Alibaba, April 2026): 35B / 3B active, 262K native context, Apache 2.0 license.
• Gemma 4 (Google, April 2026): 26B-A4B MoE (~3.8B active), 14 GB Q4 footprint, ~85 tok/s on consumer hardware, open weights under the Gemma license.

The performance threshold where local inference becomes indistinguishable from a cloud API is roughly 30 tokens/second, but the path to clearing it depends sharply on architecture. Memory bandwidth is the limiting factor for token generation, not memory capacity. RTX Spark and DGX Spark are unified-memory designs running LPDDR5X at roughly 250 to 300 GB/s; Strix Halo runs LPDDR5X at ~256 GB/s; for comparison, the dedicated-VRAM RTX 5090 hits 1.79 TB/s on GDDR7, and an H100 hits ~3.35 TB/s on HBM3. That is a five-to-seven-times bandwidth gap between the consumer unified-memory tier and the dedicated-VRAM tier, and for dense models (where every token requires streaming the full weight set) the gap maps almost directly to throughput. The unified-memory tier can fit large dense models but it cannot run them at usable speed: a 70B-parameter dense model at Q4 weight (~40 GB) runs at roughly 5 to 10 tok/s on these platforms, well below the 30 tok/s threshold.

What the unified-memory tier does run well is sparse-activation Mixture-of-Experts (MoE) models, because each token only reads the small subset of expert weights actually activated. Qwen3-Coder 30B-A3B has 30B total but only ~3B active parameters per token, so the bandwidth requirement per token is roughly one-tenth what a dense 30B model would need; Strix Halo posts 70 to 100 tok/s on this workload. Gemma 4 (26B-A4B MoE, ~3.8B active) hits ~85 tok/s on the same hardware. DeepSeek V4 Flash (284B total / 13B active) and Llama 4 Scout (109B total / 17B active) follow the same pattern. The local-inference erosion of cloud lock-in is, more precisely, a MoE-class erosion, not a uniform erosion. NVIDIA's 120B-parameter laptop claim is achievable specifically because the targeted architectures are MoE; the equivalent dense 120B model would not be a realistic interactive workload on a 300 GB/s laptop. The cloud-quality experience now ships on a desk for under $5,000 for the MoE family. For dense 70B and larger workloads the cloud is still the better venue.

The software stack that turns this hardware into usable consumer and commercial AI tools is no longer the constraint either. Ollama (Apache 2.0) provides one-line local model deployment with a REST API compatible with the OpenAI inference contract, which means a developer pointing their existing OpenAI client at http://localhost:11434 can switch from paying per token to running free locally with no other code change. LM Studio ships the same capability as a polished desktop GUI for non-technical users. llama.cpp and its derivatives power the bulk of these stacks and now ship hardware backends for CUDA, Metal, Vulkan, ROCm, and ARM, covering every consumer silicon family by default. On the commercial side, Microsoft's Copilot+ PC program shipped a Windows-native on-device AI runtime in 2024-2025 across Qualcomm, Intel, and AMD Copilot+ silicon, with Apple Intelligence doing the same on M-series Macs and recent iPhones. The IDE-and-editor tier (Cursor, Continue, Aider, Zed) all support local-model backends in addition to the cloud APIs, so the developer-tools layer that drove enterprise OpenAI dependency is portable to the local stack the day the local hardware shows up. The pattern is the same across every category: the runtime, the tooling, and the model weights are all available under permissive licenses; the only thing the hyperscaler-cloud model was bottlenecking was the silicon, and the silicon arrived in 2026.

The most leveraged single workflow inside this transition is code generation, both because it is the highest-revenue AI use-case currently funding cloud capex and because it is the cleanest published benchmark battleground. Raw model capability has converged: Qwen3-Coder 30B-A3B, DeepSeek-Coder V3, GLM-4.6, and Llama 4 Scout now match or exceed GPT-4-class scores on HumanEval, MBPP, and increasingly on SWE-bench Verified at sizes that fit in 32 to 80 GB of unified memory, well inside Strix Halo / DGX Spark / RTX Spark envelopes. But raw capability is not what makes code generation useful in production. What makes it useful is the harness: a structured framework of skills and tools wrapping the model (Aider, Continue, Cursor-style file-level diff orchestration; sub-agent decomposition into Plan / Explore / Edit roles; MCP servers that expose the existing dev toolchain as callable functions); efficient context storage that extends effective working memory far beyond the native token window (vector embeddings of the codebase; AST-indexed and symbol-indexed databases keyed to file and line; semantic-chunked RAG that retrieves only the relevant 5,000 tokens out of a 5,000,000-line repo); and tight iteration loops that feed test results, type-check errors, lint output, and runtime stack traces back into the next turn so the model converges on a working diff rather than emitting plausible but broken code in one shot. All three of these layers are open-source and portable; none of them require a hyperscaler cloud. A developer running Aider plus Qwen3-Coder 30B plus a codebase-embedding RAG layer on a $4,000 DGX Spark in mid-2026 has, for the in-house code-modification workflow that drives the bulk of enterprise developer-tools spend, a stack that eighteen months earlier required Claude Sonnet or GPT-4 plus Cursor plus per-token cloud billing. The capability differential between the local stack and the cloud frontier is, for this specific workflow, no longer the moat the cloud capex commitment assumed.

The implication for the hyperscaler-cloud capex bet is direct, with the architectural qualification just made. Every consumer-tier laptop or mini-PC that ships with MoE-capable unified-memory hardware in 2026 is a future seat that the hyperscaler-cloud AI revenue trajectory above does not get for the MoE-served portion of demand. Every commercial workflow that ports to a local-first MoE runtime is a recurring per-token billing relationship that does not renew at the same magnitude. The local stack is not a complete substitute for cloud AI (frontier-only dense workloads at 70B+, bursty load that cannot amortize against a fixed local footprint, regulated industries that need managed compliance, anything where memory bandwidth or training matters all still belong to the cloud), but it does not need to be a complete substitute to shrink the addressable market that the $700B-by-2027 trajectory implicitly assumes. It needs to be a credible alternative for the MoE-architected, high-volume, latency-tolerant, stable-workload bulk of enterprise inference demand, which is a meaningful share and growing as more model labs publish MoE-default open weights.

The strategic implication is sharp. NVIDIA's revenue is hedged across every layer: cloud (H100 / B200 / GB200), data-center workstations (DGX), prosumer (DGX Spark), consumer desktop (RTX 50-series), and laptop chips. The hyperscalers are not. Their capex commitment was placed on the premise that enterprises and consumers cannot run useful AI locally and must therefore rent inference from their clouds. As that premise erodes:

• Enterprises with stable, MoE-served inference workloads gain a credible path to bring AI in-house. An organization paying $X million per year to OpenAI / Anthropic / Bedrock for workloads that fit on Qwen3-Coder / DeepSeek V4 / Llama 4 Scout-class MoE models can buy a fleet of DGX Sparks at ~$5K each and amortize the compute in months. Dense 70B+ workloads still require the cloud (or much more expensive multi-GPU local hardware) because of the memory-bandwidth gap.
• Downstream small and mid-market businesses gain the ability to deploy AI without becoming permanent cloud customers, the precondition for AI to be a productivity tool rather than a rent-extraction layer.
• The cloud-AI providers' addressable market shrinks to the use cases where renting still makes economic sense (bursty load, low-latency frontier-model requirements, regulated industries that need managed compliance), which is a much smaller market than the universal-cloud-lock-in story the capex committed assumes.

The bet on the cloud side is therefore not just on AI's commercial success at scale, it is on the cloud lock-in holding long enough to amortize the committed capex before consumer hardware closes the capability gap further. The single company with the most insight into how fast that gap is closing is also the one selling the products doing the closing. NVIDIA's interests align with whichever side wins; the hyperscalers' do not.

The implicit backstop: why the bet looks rational from the inside

The capex numbers invite the obvious question: how is this rational? Given the capital intensity and the 2027 trajectory toward $1 trillion in annual industry capex, the equity-IRR math is brutal. The math closes only under one assumption: that the bag will not actually be held by the people who placed the bet.

The federal pattern is documented across five major interventions in eighteen years:

• 2008: TARP authorized $700 billion to recapitalize banks; AIG received $182 billion in lifelines after its CDS book imploded; the major US automakers (GM, Chrysler) received separate auto-industry bailouts.
• September 2019: the overnight repo market seized up. Rates spiked from roughly 2.2% to 10% as collateral demand overwhelmed dealer balance sheets. The Fed restarted balance-sheet expansion (the "not QE" episode that became full QE within months) and eventually formalized the Standing Repo Facility (SRF) in July 2021 as a permanent on-demand backstop for the Treasury-collateral repo market, a tool that did not exist as standing infrastructure before the 2019 episode required it.
• 2020: the CARES Act authorized $2.2 trillion, with roughly $500 billion earmarked for airlines and large-employer support; PPP added approximately $800 billion for small business; the Federal Reserve's Primary and Secondary Market Corporate Credit Facilities directly purchased investment-grade corporate debt for the first time in the central bank's history.
• 2023: the SVB and Signature rescues invoked the FDIC's systemic risk exception (12 March 2023) to protect uninsured depositors above the $250,000 cap, effectively making the tech-VC ecosystem whole with public balance-sheet exposure. First Republic (1 May 2023) was separately resolved via FDIC purchase-and-assumption with JPMorgan, without invocation of the SRE.
• 2024-2026: NYCB-class regional banks have received quiet capital infusions and forbearance on commercial real-estate concentration; the Standing Repo Facility has been repeatedly expanded and made operationally permanent. As established by the FOMC on 28 July 2021 (Federal Reserve Board press release, Statement Regarding Repurchase Agreement Arrangements), the SRF was authorized at a $500 billion maximum aggregate operation size, against Treasuries, agency debt, and agency MBS, with counterparties initially limited to primary dealers and explicit FOMC-chair discretion to temporarily raise the cap. Depository-institution access was added on 1 October 2021 and has since expanded; daily operations now run continuously and function as a backstop for the primary-dealer repo market that funds Treasury collateral and short-term dollar liquidity globally. FHLB liquidity windows have remained open beyond their stated purpose.

The pattern is consistent: when the failure mode threatens systemically-important financial institutions or major employers, the government socializes the loss. Dodd-Frank's "living wills" regime nominally curbs this, but the 2023 SVB exception in particular established that the policy is exception-based, not rule-based, which is to say, it is the rule.

For the hyperscaler-and-LP coalition placing the 2026 AI-infrastructure bet, the implicit calculation has a clean shape:

• The bet's upside accrues to equity holders directly.
• The bet's downside, if it triggers, propagates through bondholder pension funds (Siphon 4), bank Tier-1 capital concentrated in commercial real estate (the demand-side of Siphon 5), regional banks with CRE exposure, municipal tax bases (Siphon 3), and the index funds that quietly rebalanced retail investors into the position (Siphon 4).
• Each of those failure modes individually meets the "systemically important" threshold that triggers historical bailout precedent.
• A coordinated AI-capex unwind would meet that threshold overwhelmingly.

Under that calculation the capex is not irrational. It is a high-confidence bet not on AI's commercial success but on the federal government's documented pattern of socializing losses when those losses concentrate in systemically-important institutions and large-employer balance sheets. The Goldman / MIT skepticism about AI commercial ROI does not invalidate the bet; it just identifies which leg of the bet pays out. If AI delivers on its commercial thesis, the equity holders win the productivity capture. If it does not, the precedent says the equity holders are still made whole through the rescue mechanism, and the realized cost lands on ratepayers, communities, taxpayers, passive investors, and other businesses.

Private gains, socialized losses, priced into the IRR ex ante.

Who pays if the bet fails is not hypothetical. The same pattern has produced public-balance-sheet exposure five times in eighteen years. The wealth transfer that pays for the bet before a failure event arrives runs through five channels; the bailout precedent is how the payment is finished if one does.

The explicit subsidy: federal capital and federal lands

The implicit backstop above is post-hoc: the federal balance sheet appears after a loss event to socialize the writedown. Beginning in early 2025 the federal government also opened a parallel ex ante channel that subsidizes the buildout directly, in two named instruments.

Stargate, announced from the White House on 21 January 2025, committed an aggregate up-to-$500 billion to AI infrastructure as a public-private joint venture between SoftBank, OpenAI, Oracle, and the UAE-backed MGX, with an initial $100 billion committed by SoftBank. The federal posture was endorsement and coordination, not equity, but the announcement itself moved the cost-of-capital math for every other hyperscaler bond issuance and equity raise that quarter. The structural read is the one that matters: a US president publicly underwriting a single industry's capex trajectory is the strongest possible signal that the implicit backstop applies if the bet sours.

Executive Order 14318, "Accelerating Federal Permitting of Data Center Infrastructure" signed 23 July 2025, did the operational follow-through. It directs the federal government to make federal land available for data-center siting, instructs the Department of the Interior to identify Brownfield and Superfund sites for data-center development, and creates a "qualifying project" threshold of $500 million in capital expenditures OR more than 100 megawatts of additional electric load OR a national-security designation, below which the accelerated permitting does not apply. The threshold is, by construction, a hyperscaler-only gate. A regional bank cannot reach $500M of data-center capex; a hospital cannot peak at +100 MW of new load. The order's faster permitting is structurally available only to the largest issuers.

On 24 July 2025, one day after EO 14318 was signed, DOE announced four federal sites moving forward with solicitations for private-sector data-center and energy-generation partners: Idaho National Laboratory, Oak Ridge Reservation, Paducah Gaseous Diffusion Plant, and Savannah River Site. Three of the four are former nuclear-weapons-complex or uranium-enrichment sites with existing grid capacity, existing security clearances, and federal control of the surrounding land. They are precisely the sites where a hyperscaler can bypass the 4-to-7-year utility-interconnection queue that currently constrains every other large commercial or industrial buyer in markets like Northern Virginia, Phoenix, and Dallas. A 100-megawatt commercial customer in Northern Virginia cannot get connected before 2030 to 2033; a hyperscaler at Oak Ridge can.

Stargate and EO 14318 do not create the wealth-transfer channels the rest of this piece documents; those channels were already running before January 2025. What the federal instruments do is lower the marginal cost of operating them by removing two of the largest frictions, permitting delay and grid-interconnection queueing, in a way that is by construction only available to hyperscaler-scale issuers. The federal subsidy is not money on the table, it is the removal of barriers to the existing transfer.

The return loop: do the recipients pay back in?

The fair test of public benefit is not whether a project employs anyone or generates any tax revenue. It is whether the recipients of public capital, public land, ratepayer-subsidized power, state abatements, and federal permitting acceleration close the loop by paying meaningful federal income tax against the profits the buildout produces. The 2025 books, the first full year of the Trump-administration "One Big Beautiful Bill Act" (OBBBA, signed July 2025) plus the renewed federal permitting posture, are unambiguous on this. They don't.

Four of the most-leveraged-to-AI public companies, Amazon, Meta, Alphabet, and Tesla, collectively reported roughly $315 billion in U.S. pretax profit for 2025 and collectively paid approximately 4.9% of that in federal corporate income tax (Institute on Taxation and Economic Policy analysis of 10-K disclosures, "Four Big Tech Companies Avoid $51 Billion in Taxes in Wake of OBBBA"). The statutory federal corporate rate is 21%. The Inflation Reduction Act's Corporate Alternative Minimum Tax (CAMT), enacted in 2022 as the explicit fix for book-versus-taxable-income divergence at firms with average annual financial-statement income above $1 billion, is on paper a 15% floor. Both rates are higher, by a multiple, than what these four firms actually paid.

The per-company numbers, from the 10-K disclosures and follow-on press coverage:

Company	2024 federal tax (US)	2025 federal tax (US)	2025 effective fed rate	Notable mechanism
Amazon	~$9 billion	~$1.2 billion (current)	~1.4% of $89.5B pretax US per ITEP	OBBBA bonus depreciation + R&D expensing applied to data-center capex; ~14K corporate cuts late 2025 + further rounds through 2026 (~30K cumulative)
Meta	(higher base)	3.5% effective federal rate	3.5%	CFO Susan Li in earnings call: "substantial cash tax savings from the new US tax laws, given the significant investments we're making in infrastructure and R&D"
Alphabet	$21.1B (fed + state combined)	$13.8B (fed + state combined)	~16.8% effective	Highest of the four; AI capex deductions still drove a $7.3B year-over-year drop
Tesla	(variable)	$0 on $5.7B US income	0% effective	Per ITEP, "Tesla Reported Zero Federal Income Tax on $5.7 Billion of U.S. Income in 2025"

Aggregate effect per ITEP: roughly $51 billion in 2025 federal tax avoidance across the four companies relative to a normalized statutory burden. That is more than the entire annual CHIPS Act semiconductor-manufacturing appropriation ($52.7B over five years), avoided in a single tax year by four firms.

The mechanism is no longer hidden. AI-infrastructure capex, GPUs, transformers, switchgear, building shells, networking, qualifies as either bonus-depreciable property or capitalized R&D under OBBBA, and OBBBA restored the immediate-expensing treatment that the 2017 Tax Cuts and Jobs Act had begun phasing out. The result is mechanical: a dollar spent on a Blackwell GPU or a Wisconsin data-center pad becomes a dollar of immediate deduction against federal taxable income. Meta's CFO said this on a public earnings call, in those words. The more the firm spends on AI infrastructure, the less it owes in federal income tax. The same capex that earlier sections of this piece document as drawn from ratepayers, communities, state taxpayers, and passive investors is also what shields the firm spending it from federal income tax against the resulting profits.

There is also a transfer-pricing legacy issue that bears mention because it is the largest open enforcement action against a hyperscaler in modern IRS history. Microsoft received Notices of Proposed Adjustment from the IRS in October 2023 seeking $28.9 billion plus penalties and interest in back taxes for tax years 2004 to 2013, related to a cost-sharing transfer-pricing arrangement with a Puerto Rico affiliate that the IRS determined to be "illusory in nature, serving no material economic purpose except to shift income to Puerto Rico" (IRS NOPA quoted in Microsoft Form 8-K, October 2023; Bloomberg Tax). Microsoft has reserved against the dispute and is contesting through administrative appeals and, if necessary, court, a process expected to take years. The Microsoft case is not 2025-vintage and not OBBBA-driven, but it is the structural backdrop: the prior decade of hyperscaler federal tax positions are, in at least one large public instance, in active dispute as not-acted-in-good-faith profit shifting.

The combined picture: the federal government opened explicit subsidy channels (Stargate, EO 14318, DOE federal-land siting) and lowered effective federal corporate tax on the recipients to single digits through OBBBA's capex-expensing posture, while CAMT's nominal 15% floor proved porous enough that 2025 effective rates ran 3% to 5% at the most-leveraged firms. The flow of public resource toward the hyperscalers, ratepayer cross-subsidy, state abatements, federal land, ratepayer power, accelerated permitting, capex-deductibility, met by federal tax payment back in the single-digit-percent range in 2025. The return loop is open at one end and effectively closed at the other.

Siphon 1: utility ratepayers

The most visible channel is your electricity bill.

In 2026, data centers account for roughly 40% of all electricity demand growth in the United States (Goldman Sachs Global Investment Research, "Generational Growth: AI, Data Centers, and the Coming US Power Surge"). Goldman and CNBC reporting on the same dataset project household electricity prices will rise an additional 6% through 2027 above the underlying trend, the data-center-attributable delta. Lower-income households, where electricity is a larger share of spending, take the largest hit.

The mechanism is straightforward. Utilities must upgrade transmission, distribution, and generation to accommodate the new load. Under the standard rate-setting model, those upgrade costs get socialized across the entire ratepayer base. A retiree in suburban Columbus, Ohio pays a portion of the upgrade bill for the AEP transmission lines that feed the next Meta or Google campus an hour away, even though the retiree's load has not changed and they will not benefit from the new capacity.

The same model has been extended to water. Texas data centers consumed 49 billion gallons in 2025 and are on track to consume up to 399 billion gallons by 2030, equivalent to drawing Lake Mead down by 16 feet per year (Lincoln Institute of Land Policy, "Data Drain: The Land and Water Impacts of the AI Boom"; Texas Water Development Board projection cited in Fortune coverage of rural-county aquifer impacts). Nearly two-thirds of new U.S. data centers since 2022 have been built in high-water-stress areas: California, Arizona, Texas.

The small countercurrents, AEP Ohio's tariff requiring developers to cover 85% of requested demand as minimum charges, Maine's LD 307 20-MW moratorium that passed the legislature in 2026 before being vetoed by Governor Mills (override failed 29 April 2026), the voluntary March 2026 White House Ratepayer Protection Pledge, are working in narrow jurisdictions but vastly outpaced by the buildout. For everyone outside those few jurisdictions, socialized cost recovery remains the default.

Siphon 2: adjacent communities

The second channel is the most visceral and the least monetized. It runs through the air, the soil, the water table, the noise floor, and the value of the homes within sight of the fence line.

Property values. Properties 200 feet from a hyperscale data center sold at a 15-18% discount; properties 1.1 miles away showed a 0-2% premium. The wealth depression is sharply concentrated at the fence line.

Northern Virginia, home to roughly 300 facilities, has nearly one-third of its data centers within 200 feet of residentially zoned property, about 100 hyperscale facilities sitting close enough to depress the wealth of the families nearest to them. That is wealth households never agreed to lend to the AI buildout.

Agricultural conversion. A substantial share of new U.S. data-center siting is occurring on former farmland in rural areas. The American Farmland Trust ("Farms Under Threat" series) counts roughly 2,000 acres of farmland lost per day to non-agricultural use; USDA data shows roughly 75 million acres of farmland lost between 1997 and 2022. The conversion is permanent: once farmland is encumbered with a 200-megawatt substation and a 50-acre cooling-plant slab, it does not return to production.

Air pollution. Virginia regulators have approved 1,937 non-emergency diesel generators for data-center use (Virginia Mercury and Sierra Club tracking of Virginia Department of Environmental Quality permits). Amazon's permit for its Canton, Mississippi facility estimates 240 tons of NOx per year, roughly nine times higher than the nearby Nissan manufacturing plant (Mississippi Free Press).

Diesel generators emit 200 to 600 times more nitrogen oxides than natural gas plants, plus PM2.5 and ozone precursors. The public-health literature on chronic PM2.5 exposure, asthma, cardiovascular disease, premature mortality, is unambiguous. California SB 978, introduced February 2026, would ban new diesel backup installations in high-pollution zones. It is one bill in one state.

Water contamination. Modern liquid-cooled and immersion-cooled data centers use per- and polyfluoroalkyl substances (PFAS), "forever chemicals", in two-phase immersion cooling and two-phase direct-to-chip cooling. PFAS are also present in cleaning agents, fire-suppression systems, and refrigerants throughout the facility. Virginia does not currently require PFAS testing of data center discharge water. The federal EPA does not require it either. The Toxic Substances Control Act PFAS reporting deadline has been pushed to January 31, 2027.

Noise. Prince William County, Virginia records routinely exceeding 60 decibels at residential property lines, well above WHO nighttime guidelines. Documented health effects on near-neighbors include sleep disturbance, hypertension, cardiovascular stress, headaches, vertigo, nausea, and ear pain. The Frontiers in Climate journal described the cumulative Virginia effect as a "health earthquake" in April 2026.

Property value depression, agricultural land loss, NOx, PFAS, and noise are each priced at zero in the developer's pro forma. They are paid for, in real dollars and real health, by the people who happen to live nearby.

Siphon 3: taxpayers

The third channel runs through state and local budgets, and it is enormous. According to Good Jobs First, three states are losing $1 billion or more per year in foregone tax revenue from data center incentives.

The Texas figure has grown almost eightfold since 2023, from a $130 million Comptroller projection (Feb 2023) to $1 billion in FY2025 (Texas Tribune; Texas Comptroller revised estimate). Wisconsin's four certified projects, Microsoft in Mount Pleasant, Epic Hosting in Verona, Meta in Beaver Dam, an Oracle-related project in Port Washington, are projected to cost the state $1.5 billion in foregone sales tax upfront plus another $369 million per year once operational. For comparison, Wisconsin's annual K-12 education budget is around $7 billion. The four data center projects, once operational, will permanently subtract roughly 5% of the state's annual K-12 spending capacity.

Good Jobs First identified 14 states that fail to disclose their data center tax shortfall. Only three, Washington, Texas, and Virginia, properly disclose the losses in their Annual Comprehensive Financial Reports per GAAP. The Register's April 2026 headline summed it up: "US states can't account for datacenter tax breaks. Literally."

The justification is always the same: jobs and capital. A fully built-out hyperscaler data center employs a permanent on-site staff that varies widely with size, automation level, and operator, from a handful of people per shift at the most automated sites (industry visitors report seeing tier-3 facilities running with three on the clock) to a few dozen at larger campuses. Almost none are local hires. The trajectory is downward by design: spinning-disk replacement is automated away as flash matures, remote-hands work is consolidated, and an increasing share of on-site labor is being converted into vendor support agreements rather than direct headcount. The capital is overwhelmingly imported equipment, GPUs, transformers, switchgear, none of it fabricated in the abated jurisdiction. What remains in the community is the power consumption, the water consumption, the diesel exhaust, the land conversion, and the political cover for the next abatement.

The land point is structural and worth stating directly. A 500-megawatt hyperscaler campus typically occupies 200 to 500 acres of fenced, secured, low-traffic footprint that contributes effectively zero foot traffic to local retail, zero residents to the school district, zero shoppers to the property-tax base, and zero per-acre jobs beyond the small on-site staff already noted. Whatever the prior use of that acreage was, the opportunity cost is the same: farmland would have produced food output and agricultural-sector employment; residential would have produced a property-tax base, school enrollment, and household consumption recirculating through local businesses; retail or commercial would have produced per-acre jobs and a sales-tax base; even a single Walmart Supercenter or a regional call center would deliver materially more local economic activity per square foot than the hyperscaler. The land is not just converted, it is functionally removed from the local economy. Local officials negotiating an abatement against the promise of "jobs and capital" are also negotiating against the foregone tax base, foregone employment, and foregone consumption that any other plausible use of the same acreage would have produced.

The "build them in the desert where nobody lives" rebuttal is structurally weak for the same reason the wealth-transfer argument is structurally strong: the data center does not actually become less resource-intensive when it is far from people, it becomes more. Water still has to come from somewhere, typically a pipeline pulled from a regional aquifer or a long-distance transfer (Texas, Arizona, and Nevada projects are the clearest cases); and a 100-degree desert ambient temperature roughly doubles the cooling load relative to a 70-degree coastal site, which means doubling the water and electricity allocated to chillers and evaporative cooling. Power still has to be transmitted in from generation, with the EIA reporting average US transmission-and-distribution losses of roughly 5% of generated electricity across 2018-2022 (EIA FAQ, "How much electricity is lost in electricity transmission and distribution in the United States?"), and long-haul AC routes typically running toward the upper end of the 5%-to-7% band, with the buildout cost socialized to the broader ratepayer base. Network fiber still has to be run from the nearest exchange point, often hundreds of miles. Replacement parts, GPUs, transformers, switchgear still ship in. The small staff still has to be either relocated, housed, and serviced locally (in which case the "nobody lives there" claim is no longer true) or rotated in via long commute (in which case the desert location is supporting commuter infrastructure on the local tax base anyway). The desert sites work for the operator's permitting timeline and political risk; they do not actually reduce the externalized cost. They mostly relocate it onto a different and typically less-organized constituency.

Siphon 4: investors

The fourth channel is the most counterintuitive, because the people on the receiving end believe they are the beneficiaries.

Capital intensity, capex as a share of revenue, has reached 45 to 57% at the major hyperscalers, levels Goldman analysts (per the "Tracking Trillions" report series) describe as "historically unthinkable." Goldman further models that shortening the economic life of AI chips from five years to three, closer to actual hyperscaler-grade utilization, would increase annual depreciation expense by nearly $1 trillion across the industry, an order of magnitude that would absorb most or all of the optimistic-case revenue growth the capex commitments depend on.

There is also a quiet capital-markets dynamic that should worry passive holders. Meta, Alphabet, Amazon, and Oracle's collective weighting in the Bloomberg U.S. Corporate Investment Grade Index nearly doubled in the year ending April 2026.

Every passive investor in a broad bond fund or pension benchmark now has materially more AI-infrastructure debt exposure than they did twelve months ago. Most do not know it, and none of them affirmatively chose it. They were rebalanced into it by the math of index construction.

The retail equity exposure is even larger and even less consensual. A 401(k) holder in a target-date fund tracking the S&P 500 currently has roughly a quarter to a third of their U.S. equity exposure in the cluster of names whose capex-to-revenue ratio is, by Goldman's own description, historically unprecedented (Microsoft, Alphabet, Amazon, Meta, Oracle, and NVIDIA as the most directly leveraged; the exact share moves with quarter-end index rebalancing and recent price action, but the order-of-magnitude approximation is stable). Whether or not the AI bet pays out, the retirement security of a generation has been quietly leveraged to it.

The ouroboros: where Siphons 4 and 5 meet

The structural pattern that emerges when Siphon 4 (passive investor exposure to AI-infrastructure capex) is read against Siphon 5 (demand contraction in the office-supporting ecosystem) is that the major US private-equity multi-strategy houses are simultaneously the LP-tier capital behind the AI buildout AND the equity holders of the office CRE that AI-driven labor displacement is now hollowing out. They are on both sides of the same trade.

House	CRE exposure (illustrative)	AI / data-center exposure (illustrative)
Blackstone	BREIT and BX Real Estate (~$300B+ AUM at peak); BREIT redemption-gated through 2022-2023 over CRE concerns	QTS Realty Trust take-private ($10B incl. debt, announced June 2021, closed 31 Aug 2021; QTS Form 8-K, SEC EDGAR); AirTrunk (A$24B / ~$16B USD, announced 4 Sep 2024 with CPP Investments as co-investor; Blackstone press release, "Blackstone Announces Agreement to Acquire AirTrunk in a A$24B Transaction"); AI-startup LP positions across BCP funds
Brookfield	Manhattan West, Brookfield Place, large global office book; Brookfield Property Partners write-downs through 2024	Brookfield Renewable powering AI data-center buildouts; AI-infrastructure fund vehicles
KKR	KREST and multi-strategy real-estate funds with significant office	Multi-strategy AI-infrastructure positions; data-center fund vehicles
Apollo	Distressed CRE acquisitions; office in special-situations book	AI compute-infrastructure plays; tactical AI equity

The mechanic in four stages: (1) PE house commits billions to AI infrastructure, thesis: productivity gain accrues to capital, equity multiplies; (2) AI deploys at enterprise scale, headcount reductions in the middle communication-relay layer (Siphon 5); (3) the substantial share of CBD floor area those roles occupied empties out; (4) the office CRE leg writes down, absorbing much or all of the AI-leg gain.

In aggregate the position can be self-cancelling, long AI infrastructure, short the wage stream that supports office CRE, without any single decision-maker recognizing it as one trade. The ouroboros is not a conspiracy; it is a coordination failure at the asset-allocation level. To the LP (pension funds, sovereign wealth funds, endowments), the AI leg and the CRE leg look like diversifying exposures with low historical correlation. They are actually negatively correlated by construction: when the AI leg pays off, the CRE leg suffers, because the AI leg is the cause of the CRE leg's suffering. The GP earns management fees on both legs. The LP eats the writedown when reality catches up.

The deeper asymmetry: labor recirculates, capital accumulates

The ouroboros is one local expression of a broader macroeconomic point that all five siphons share but that none of them name directly: this is a transfer from labor to capital, and labor and capital have radically different economic velocities.

A dollar of labor wage pays income tax, payroll tax, FICA, state withholding, a combined effective rate of typically 25-35% for middle-class earners. What remains flows almost entirely into consumption: rent or mortgage, groceries, school fees, the local doctor, the dry cleaner, the daycare, the cobbler, the Christmas list. The marginal propensity to consume (MPC) for the bottom three quintiles is near 1. That spending recirculates, it becomes second-order income for other workers and other small businesses, who in turn pay taxes and consume.

A dollar of capital return faces a materially lower effective tax rate (long-term capital gains 15-20%, step-up basis at death, trust and family-office structures that defer or eliminate). What remains flows predominantly into asset acquisition: more equity, more Class A real estate, more art, more private credit, more index funds. The marginal propensity to consume at the top of the wealth distribution is near zero. Capital does not Christmas-shop. It does not stock the grocery aisle. It does not fund the school district. It does not pay the line cook. It does not buy the dry cleaner's services. It accumulates.

When economic value flows from labor (high-velocity) to capital (low-velocity), the aggregate spending pool shrinks even when measured GDP does not fall. The five siphons all share this velocity asymmetry: utility ratepayers, adjacent communities, state taxpayers, passive investors-by-default, and other businesses are all categories whose effective MPC is high. The hyperscaler equity holders capturing the transferred value have MPC near zero.

This is the part of the office-CRE story that the bank-balance-sheet framing most cleanly misses. Banks see the loan losses. They do not see the cascade of demand destruction that follows when the wage stream supporting a dozen small businesses per office tower is permanently capitalized into one hyperscaler's net income line. The CRE writedown is the headline. The lost retail spending, lost municipal tax base, lost public services, and lost downstream small-business equity is the tail, and the tail is structurally bigger than the headline.

Siphon 5: other businesses

The fifth channel is the most economically distorting and the least often discussed. Other businesses face a two-sided squeeze: the inputs they need to operate are increasingly allocated to hyperscalers, AND the wage stream that buys their output is being hollowed out as AI displaces the white-collar work that drove demand. Input squeeze on one side, demand contraction on the other. Three fronts.

Grid capacity. Grid interconnection queues in the major data-center markets, Northern Virginia, Phoenix, Dallas, are now running 4 to 7 years. A new commercial or industrial customer that submits an interconnection request in May 2026 in Northern Virginia cannot realistically expect to draw utility power before 2030 to 2033. Between 30% and 50% of large data centers scheduled to open in 2026 will be delayed or cancelled because the grid cannot deliver. High-power transformers now have 3-to-5-year lead times; switchgear is sold out through 2028. What gets crowded out is everyone else: a new manufacturing facility, a regional logistics warehouse, a CHIPS Act packaging plant, a hospital adding an MRI suite, a small data center serving a regional bank.

Hardware allocation. Microsoft, Google, Meta, and Amazon placed multi-billion-dollar forward orders for NVIDIA Blackwell GPUs in 2025, consuming most of NVIDIA's allocation through the end of 2026 and into 2027. TSMC's CoWoS packaging, the bottleneck for binding HBM stacks to GPU dies, is fully allocated through at least mid-2027. A hyperscaler will always allocate reserved capacity to a $100M enterprise customer before making on-demand inventory available to a startup. Mid-market and enterprise customers who previously bought through standard channels have been crowded out entirely. Even Chinese companies cut off from H100s and H200s have responded by stockpiling whatever they can get, distorting global pricing and pulling supply away from non-hyperscaler buyers in unrelated markets.

Demand contraction, the office-supporting ecosystem. The companion to input crowding is customer disappearance. AI's displacement target is precisely the middle communication-relay layer that occupies a substantial share of the floors in any major central business district: project managers, business analysts, junior consultants, product managers, account managers, executive assistants, internal communications, HR business partners, paralegals, junior audit and accounting. Their job is to translate executive intent into developer action, take a directive, produce specs, summarize feedback, draft documentation, coordinate calendars, run meetings, shepherd execution. That is exactly the work current large language models do well, and it is exactly the work that requires office presence to do well (you cannot coordinate other people as effectively in async chat; you cannot build relationships in a Slack thread; you cannot run a productive working session over Zoom for a group that has not first sat together). The role is simultaneously the most exposed to AI AND the most load-bearing for the office. The same people who fill the Class A floors are the people whose work LLMs can now do.

The leading edge is already visible in named restructurings. Fidelity Investments confirmed roughly 800 cuts on 8 May 2026, approximately 1% of an 80,000-person global workforce, concentrated in technology and product-delivery teams as the firm dismantles its "agile squads" structure for larger, more centralized teams (Bloomberg, May 2026). Recruiter-and-comp analyst Amanda Goodall's public tracking of internal RIF data (X, "the data I looked at shows... support hiring slowed down big time, middle layers started shrinking, engineering roles are protected") flagged the three features that matter for this thesis: support hiring slowed sharply, middle layers started shrinking, and engineering roles were explicitly protected. Fidelity also ordered roughly 25,000 employees back to 5-day in-office beginning September 2026, ending the hybrid arrangement that had been in place since the pandemic. That is the exact cut-and-recall profile a labor-substituting AI rollout produces: thinner middle layer, preserved technical core, denser physical presence for the survivors.

Fidelity is the cleanest single example, but it is part of a documented 2024-to-2026 pattern of delayering restructurings whose cut profiles map onto the same thesis. Amazon's Andy Jassy mandated a 15% increase in the ratio of individual contributors to managers by the end of Q1 2025, raising span-of-control from 6 to 8 reports per manager; Morgan Stanley estimated up to 13,800 of Amazon's 105,770 managers could exit under the guidance, executed through team consolidations and manager demotions rather than headline mass layoffs (Fortune, March 2025; Entrepreneur, October 2024). Microsoft cut roughly 6,000 roles in May 2025 and another ~9,000 in July 2025 (over 15,000 for the year), with CFO Amy Hood telling investors on the April 2025 earnings call that the firm would continue "increasing our agility by reducing layers with fewer managers" while redirecting capital toward an $80 billion AI infrastructure spend; Microsoft Customer and Partner Solutions, a sales-and-account-management division, was among the hardest hit (Bloomberg / TimeTrex / TheHRDigest coverage of the May and July 2025 RIF rounds). Meta announced ~3,600 cuts in February 2025 targeting "lowest performers" (employees with positive prior-year reviews were among those cut), then layered an additional ~8,000 cuts in May 2026 as Zuckerberg's $125-145 billion 2026 AI infrastructure guidance range ramped (CNBC, January and May 2025; The Next Web, May 2026; Fortune, April 2026). Salesforce CEO Marc Benioff said on 2 September 2025 (Logan Bartlett podcast; CNBC) that the firm had reduced support headcount "from 9,000 to about 5,000" with the difference covered by AI agents ("I need less heads"), and that customer-support cost had declined 17% over the period (Fortune; CNBC; ODSC summary).

The counterexample that proves the pattern is Klarna, which in 2024 publicly replaced ~700 customer-service agents with AI before quietly reversing course beginning May 2025 and rehiring after customer-satisfaction data deteriorated on complex interactions (CEO Sebastian Siemiatkowski to Bloomberg, May 2025: cost had been "a too predominant evaluation factor" producing "lower quality"). Klarna's reversal is a reminder that the substitution is not frictionless and that the first wave of AI-justified cuts will include some that get partially undone; it does not invalidate the structural direction, only the precise pace.

The common shape across these five firms is the same: the bulk of the cuts fall in manager, middle-layer coordinator, and customer-service roles, while engineering and the most senior decision-makers are preserved or expanded. The load-bearing tenant of the modern Class A office is being thinned by name, at the largest firms, on the public record, with the AI-infrastructure capex commitment cited explicitly in the same earnings-call language.

The wages of this layer, typically $80K to $200K, are also the salary band that drives CBD retail. The executive suite is too small to fill the lunch spots; the developer cohort is increasingly remote anyway; it is the middle communication-relay layer that buys the salad bowls, the dry-cleaned shirts, the commuter-rail passes, the after-work drinks, the holiday catering. When this layer thins, and the early 2026 enterprise rollouts are only the leading edge, the supporting-business demand goes with them.

The downstream cascade is already visible in the office-vacancy numbers. Major-CBD office vacancy in Q1 2026 sits well above the pre-pandemic 10-15% baseline.

The supporting-business ecosystem indexed to a full office tower is multi-layered and small-business-heavy: in-tower coffee carts and lunch spots, in-block dry cleaners and restaurants and dentists, in-CBD hotels and commuter parking and business-attire retail, the transit agencies whose fare-box recovery assumes Monday-Friday commuter density, the commercial cleaning / HVAC / elevator / security contractors whose unit economics assume full floors, and the architecture firms and CRE brokers whose pipelines assume periodic build-outs.

This is demand-side crowding-out, not input-side. WFH (2020-2024) was geographically redistributive, the lunch spot in the Loop lost; the lunch spot in Oak Park gained; net retail employment moved but didn't vanish. AI displacement is structurally different: the worker doesn't relocate, the worker is laid off. The wages disappear. The displaced-worker household deleverages. The Oak Park lunch spot doesn't gain to offset the Loop's loss. The capital owners who captured the productivity gain spend it on luxury goods and asset acquisition, neither of which props up the supporting-business ecosystem at any geography.

Bank exposure (CRE loans, regional banks with Tier-1 CRE concentration above 300%) is the visible end of this. The actually-exposed surface is much broader: the supporting-business owners themselves (often immigrant, often undercapitalized, no balance sheet to absorb the loss), municipal tax bases (San Francisco's projected two-year general-fund deficit grew to roughly $937 million by November 2025, with commercial-property reassessments cited as a primary driver per the SF Office of the Controller, SF Standard, and Real Deal reporting), municipal bondholders, public-sector workers whose pensions depend on CBD property-tax revenue, and the supporting-business workers themselves (line cooks, janitors, dry-cleaner counter staff). None of these show up on a hyperscaler's pro forma.

The compound effect is two-sided crowding-out: businesses are squeezed on what they need to operate (electricity, water, equipment, GPUs) AND on the customer base whose wages buy their output. The market clearing mechanism on the input side is price (power purchase agreements at premium rates, GPU rental costs surging 40% on the spot market, transformer leadtimes that push capex projects out of viable IRR windows). The market clearing mechanism on the demand side is failure (small businesses close, commercial real estate writes down, municipal services contract, residential follows). The businesses that lose this competition are precisely the ones the broader economy depends on for non-AI productivity growth and for the bottom three quintiles of employment.

The transfer, in one frame

The five siphons compound. A taxpayer in Georgia is also an electricity ratepayer in Georgia, is also a homeowner whose property sits in the noise contour of a Loudoun-style cluster, is also a 401(k) holder with hyperscaler exposure, is also the owner of a regional business waiting for grid capacity and GPU allocation. Each individual siphon, measured against its own baseline, is plausibly small enough to defend on jobs-and-growth grounds. The combined siphon, measured against the household's full economic exposure to the AI buildout, is not.

On the other side of these five channels sit the equity holders of approximately a dozen companies. The buildout produces a continuous transfer from the public to those balance sheets: from residential ratepayers via electricity; from neighbors via property value, air, water, and sleep; from state treasuries via abatements; from passive investors via index-driven exposure to historically-unprecedented capital intensity; from competing businesses via input crowding. The transfer is not hypothetical. It is happening at a measured rate.

The transfer in dollars: what is realized, what is exposure, what is tail-risk

We can put numbers on this. The honest answer separates three categories: realized 2025-2026 annual transfer (measurable today against sourced baselines); exposure (positions already taken that become a realized transfer only if the bet breaks); and tail-event additional loss (the magnitude of the realized transfer if 2027 revenue clears the verdict mark or it does not). Each category has a defensible range; none of them is a precise point estimate.

Realized 2026 annual transfer, measured per siphon:

The same numbers can be turned around and stacked as a share of the hyperscaler capex commitment itself, mapped quarter by quarter from Q1 2025 (the first full quarter of the 2025 ramp) forward. That framing answers: in each quarter, how much of the capex the Big Four reported is being indirectly contributed by the public through the five siphons, and how much is genuinely funded out of internal cash flow at market-rate cost of capital?

The two segments are growing at different paces, and that difference is the structurally important number. Corporate cash flow capex grew at an average of ~15% q/q across 2025 and printed at ~10% q/q for Q1 2026 (current CFO commentary points to acceleration from here, not deceleration; consensus 2027 projections are above $1 trillion industry-wide per CNBC, April 2026). The siphon-contributed share grew at ~22% q/q on average across 2025 and accelerated to ~29% q/q in Q1 2026 as OBBBA's depreciation provisions activated. Siphon growth is outpacing the recent capex-growth rate, and the divergence is widening rather than narrowing. If both observed paces continue, the lines cross.

The addressable-market ceiling: who can actually use this, and what would it cost to serve them

The crossover above assumes the demand side of the AI thesis keeps growing. The TAM-saturation question is whether it actually can. The honest answer puts a hard ceiling on the revenue side and a hard floor under the infrastructure cost that would be required to reach saturation. Both ceilings are well below what the 2026 capex commitment implicitly assumes.

The accessible-user pool is structurally smaller than the global population. Of roughly 8.1 billion people on the planet, at least 4 billion sit outside the operational envelope of Western-style AI consumption: they lack reliable always-on electricity, lack the persistent broadband connection that an interactive LLM session requires, are on mobile-first stacks where the cloud-AI economic model (per-token billing on long-form sessions) does not match how they use their device, or live in language and locale support that the major model labs do not yet cover well (the Mild Take's own work on locale-coverage gaps in the consumer AI stack documents this directly, codecai.net). That leaves roughly 3.5-4 billion people in markets where current Western-style AI tools are operationally usable, and within that pool, the realistically engaged user base at current pricing is roughly 1 to 1.5 billion. (ChatGPT alone reached roughly 900 million weekly active users by February 2026; Claude, Gemini, and Copilot add further unique users, with combined cross-platform unique reach above 1 billion globally including heavy overlap; the broader pool with the means and interest sits in the low-to-mid billions; the population with operational access caps at ~4B.)

Saturation revenue at current per-user economics. Combined cluster AI revenue at Q1 2026 sits at roughly $85-100 billion annualized (Azure AI + AWS Bedrock + Google Cloud AI portion + OpenAI + Anthropic + the long tail) against a current engaged-user pool of roughly 800M to 1B. That implies a current revenue-per-engaged-user of approximately $85-125 per year. If the engaged-user base expands all the way to the operational ceiling of 4 billion at constant per-user economics, the saturation revenue is roughly $340 to $500 billion per year. That is a wide band, and it is below Goldman's $700 billion 2027 verdict mark even at the high end. There is no headroom for the per-user economics to compress (the local-stack erosion documented earlier in the hardware-moat section pushes the per-user number down, not up); there is no obvious lever to raise the per-user number that does not collide with the same enterprise ROI ceiling already documented; and the engaged-user pool cannot expand past 4 billion without solving the structural access problems (electricity reliability, broadband penetration, locale coverage) that the model labs do not have a credible path to solving on their own.

Now the cost of actually serving that 4-billion-user saturation. Current US data-center electricity demand sits at roughly 150-200 terawatt-hours per year and is on track to roughly double by 2028 just to serve the current ~1B engaged-user pool. A 4x scale-out of the engaged user base would require building out roughly 4x the dedicated AI-compute electricity supply globally, with a similarly proportionate scale-out of fiber, water, and data-center floor space. Order-of-magnitude:

• Power generation: 200 to 400 GW of new dedicated AI-compute generation capacity globally to support full TAM saturation. At post-2024 build costs of $2.2-$2.5 billion per GW for combined-cycle natural gas (Utility Dive / NextEra), $1-$1.4 billion per GW for utility-scale solar (intermittent), and $10-$15 billion per GW for new nuclear (Vogtle came in at ~$15.7B/GW), the all-in generation build is in the $500 billion to $2 trillion range.
• Data-center compute capacity: 200 to 400 GW of new IT-load capacity at roughly $25 to $50 million per MW (the post-2024 hyperscaler build cost), or $5 to $20 trillion in new data-center construction.
• Transmission, cooling, water, fiber: typically 20 to 30% of generation + DC cost for the supporting infrastructure, or another $1 to $5 trillion.

The total infrastructure build to actually serve the TAM-saturation revenue is therefore on the order of $6.5 to $26.5 trillion, with a central estimate near $12 to $15 trillion, over a 5-to-10-year build-out window. That is more than the cumulative 2026-2031 AI capex Goldman's broader aggregate already forecasts ($7.6 trillion). Most of the headline 2026 capex commitment is the down-payment on this build, not the build itself.

The math doesn't close. Even under generous assumptions, saturation revenue is in the $340-500 billion per year range, and the cumulative capex to reach saturation is in the $6.5-26.5 trillion range with a central estimate near $12-15T. Simple payback against the range is 13 to 78 years (central estimate roughly 28 to 36 years), against a chip-economic-life that Goldman's own modeling assumes at three years and a depreciation cliff that compounds quarterly. The dollar-on-dollar return at full saturation, before accounting for operating costs, is roughly 1.3 to 7.7 cents per dollar invested per year (central estimate ~3 cents). That is not a hyperscaler-equity-return profile. It is a regulated-utility return profile, without the regulated-utility pricing protections.

What that means for the article's thesis: the capex commitment is not just betting on AI commercial success in a fixed-TAM market. It is betting on simultaneously expanding the TAM, holding per-user pricing, and absorbing a multi-trillion-dollar additional infrastructure build to actually serve that expanded TAM. Each of those three legs has to clear independently. If any one of them fails (most likely: the TAM does not expand because the access barriers are structural; or per-user pricing collapses because local inference undercuts the per-token cloud billing model), the realized 2025-2026 wealth transfer through the five siphons becomes not the down-payment on a productive industry but the entirety of what the public ever gets out of the trade.

The same numbers can also be turned around and stacked as a share of the 2026 annual hyperscaler capex commitment itself.

Underlying exposure (positions already taken, distinct from the annual flow shown above):

• 2026 hyperscaler capex commitment: ~$725 billion placed by the Big Four; the broader AI-infrastructure aggregate cleared ~$765 billion. If the AI commercial thesis pays out, this is profitable investment, not transfer. If it does not, this is the principal that absorbs the writedown.
• Passive equity holders: a 401(k) holder in an S&P 500 target-date fund carries roughly a quarter to a third of their US equity exposure in the cluster (MSFT, GOOGL, AMZN, META, ORCL, NVDA), or roughly $2.5-3.3 trillion across US retirement assets at index-driven concentration risk. This is a stock figure (cumulative position) and is shown here separately from the annual flow per-channel chart above; the annual incremental bond-side risk-shift ($80-120B/year, already included in the per-channel chart) is the flow piece.
• The cumulative committed and at-risk capital across capex + accumulated bond exposure + equity-concentration risk lands well into the multi-trillion-dollar range by end-2027.

Tail-event additional transfer (realized only if the 2027 verdict misses):

If 2027 hyperscaler AI revenue clears Goldman's $700 billion mark, the realized transfer to date stays roughly at the 2026 annual rate compounded forward, and the capex is amortized profitably. If 2027 revenue clears only the $300 billion range (the slow-case path implied by sustained sub-25% quarterly growth), the writedowns produce a second wave of realized transfer: equity impairments of roughly $300-600 billion at the hyperscaler level absorbed by passive investors and 401(k) holders; CRE writedowns on the office-supporting ecosystem of roughly $200-500 billion absorbed by regional-bank Tier-1 capital, REIT holders, and municipal tax bases; regional-bank losses of roughly $50-150 billion triggering bailout precedent and socializing further loss onto the federal balance sheet. The tail-event additional transfer in this scenario sits in the $500 billion to $1.5 trillion band, on top of the realized 2025-2026 annual rate.

To whom does the transfer flow:

The recipients concentrate at the equity layer of approximately a dozen companies plus the LP-tier asset-management houses. NVIDIA captures the hardware margin (Q1 FY27 run-rate implied ~$230B annualized net income from a single firm). Microsoft, Alphabet, Amazon, Meta, and Oracle capture the cloud-revenue ramp and, more importantly for the realized-transfer math, the OBBBA-driven federal tax avoidance that retains profit on corporate balance sheets rather than recirculating it through federal revenue. OpenAI and Anthropic capture the model-lab revenue (with OpenAI still loss-making). Tesla is included in the ITEP four-firm aggregate. Blackstone, Brookfield, KKR, and Apollo capture GP fees on the AI-infrastructure fund vehicles and the equity-return upside on the AI leg of the ouroboros trade. The equity holders of these entities skew sharply toward the top wealth decile: founder/executive concentration plus passive-index accumulation in the highest-MPC-zero tier. The transfer flows from constituencies with marginal propensity to consume near 1 (ratepayer households, adjacent-community homeowners, state taxpayers, the small-business owners hit on the input side) into balance sheets whose marginal propensity to consume is near 0 (founder-and-executive equity concentration, family-office and trust-held passive equity). That is the labor-to-capital velocity asymmetry expressed as a dollar flow.

The honest scope of what we cannot yet measure: the demand-contraction component of Siphon 5 (lost retail sales, lost CBD foot traffic, lost small-business equity downstream of office vacancy) is not yet attributable to AI displacement cleanly enough to put a number on; it will become measurable as the named-restructuring wave compounds through 2027. The grid-capacity opportunity cost (the manufacturing facilities, hospitals, and CHIPS Act packaging plants delayed by the interconnection queue) is real but not yet booked. The PFAS and air-quality health costs are real but await the January 2027 TSCA disclosure cycle to be quantified. These uncounted siphons would shift the realized 2026 annual midpoint upward, not downward.

What's actually working

The defenses exist in fragments. AEP Ohio's data center tariff is a working model for utility cost allocation. Maine's moratorium is a working model for state-level constraint. The Good Jobs First reporting is a working model for tax-transparency advocacy. California SB 978 is a working model for diesel-generator emissions reform. Tucson's water-conservation ordinance is a working model for community-level water defense. Virginia class-action noise litigation is a working model for community-level health defense. The bondholder concentration data is publicly available; institutional asset managers can choose to underweight. The grid-capacity allocation problem has known solutions, capacity auctions, tiered queue priority, lifeline reservations for non-hyperscaler load, that have been used in other contexts.

What is missing is consolidation. There is no jurisdiction in the United States where all five channels are being actively managed against the same baseline. There is no federal framework that requires hyperscalers to disclose their externalized costs on a consistent schema. There is no regulatory body whose mandate includes "what is the net effect of this buildout on the rest of the economy."

What to watch

Five signals will indicate whether this dynamic resolves or accelerates.

First, state-level disclosure. If more than half of U.S. states begin properly disclosing their data center tax expenditures in their annual financial reports, the GAAP standard the three transparent states already meet, that is evidence the political economy is catching up to the fiscal reality.

Second, utility tariff structure. If the AEP Ohio model becomes the default across PJM and ERCOT, residential ratepayers are insulated from the worst of the cost-shifting.

Third, community-impact litigation outcomes. The Virginia, New Jersey, and Mississippi class actions over noise, air, and water externalities are the early test cases. If plaintiffs win meaningful injunctive relief, the precedent forces developers to price these costs into their pro forma.

Fourth, federal PFAS disclosure. The TSCA reporting deadline of January 31, 2027 will produce the first comprehensive industrial PFAS inventory. If data centers feature prominently, the regulatory response could be significant.

Fifth, the capex-to-revenue gap. If 2027 AI revenue at the hyperscaler level clears $700 billion, the bet pays out and the externalized costs were a tolerable bridge. If it clears only $300 billion against $1 trillion in capex, the writedowns that follow will produce the largest single transfer of value from passive investors to selling executives in U.S. corporate history, and the externalized costs will have been paid for, in retrospect, nothing.

The trajectory the AI cluster has to ride is concrete enough to put numbers on. As of Q1 to Q2 2026, combined AI-attributable revenue across the major hyperscaler-cloud AI segments and the model labs (Microsoft Azure AI ~$37B annualized per Q1 FY26 disclosure, up 123% YoY; AWS Bedrock and GenAI run-rate >$15B per AWS CEO commentary; Google Cloud AI run-rate inside a $20B-quarter / $80B-run-rate Cloud business growing 63% YoY; OpenAI $24B annualized at $2B/month per the company; Anthropic ~$47B annualized as of late May 2026, up from $30B in April and $9B at year-end 2025) sits at roughly $130 to $160 billion in combined AI-attributable run-rate, with care not to double-count the model-lab revenue that flows through hyperscaler-cloud line items. For the cluster to clear what would amount to a $700-billion 2027 revenue mark (this is the article's derivation from Goldman's capex-justification framing, not a Goldman-stated revenue target), that base needs to compound at roughly 30% per quarter, sustained, for six consecutive quarters. The on-track waypoints fall out of the math:

The clean read: if the segment-level growth rates print above 30% quarter-over-quarter sustained for two or more quarters, the bet is tracking. If they print below 20% quarter-over-quarter for two consecutive quarters in any of the three large cloud-AI lines or in the two model labs, the bet is breaking. The verdict on which of these obtains will be formal within roughly 24 months, when 2027 hyperscaler revenue prints against the capex run-rate, but the leading indicators that telegraph the answer will arrive much sooner and in public. Quarterly hyperscaler capex guidance is the cleanest near-term signal: an unexplained downward revision at any one of the four would imply the thesis is breaking internally before it breaks publicly. Hyperscaler IG-bond credit spreads relative to the rest of the IG universe are a market-priced read on the same question, refreshed daily. OpenAI and Anthropic monthly revenue ramps, when they leak or get filed, narrow the commercial-success-or-failure band faster than the 24-month formal verdict. CBD office vacancy and named-restructuring announcements expand or contract the demand-contraction thesis quarterly. NVIDIA's data-center revenue mix shift between cloud GPU shipments and DGX-Spark-class on-premise hardware is a direct read on whether the cloud lock-in is holding. McKinsey, BCG, and Gartner enterprise-AI surveys publish on roughly six-month cycles. By the time the formal 2027 numbers print, the answer will already be discounted into bond spreads, capex guidance, and quarterly enterprise-software backlog.

In the meantime, the siphons continue running. The official story is that the AI boom is being financed by venture capital and corporate cash flow. The unofficial story is that it is being financed, in large part, by the people who never agreed to the bet, and who have no way to opt out.

Source-canon corroboration

The article's claims map to The Mild Take's curated resource canon as follows. The canon supplies the site's preferred independent angle on each claim type, used to triangulate the specific named citations below.

Siphon	Claim type	Canon resource	Methodology angle
1, utility ratepayers	Macro electricity-cost effects, inflation pass-through	IMF (Article IV staff reports; WEO), FiveThirtyEight data (energy datasets)	Goldman / CNBC are US ownership-tilted; IMF is the site's preferred independent macro reference
2, adjacent communities	Facility permits, NOx, water discharge, noise, property	Bellingcat (OSINT verification of permits, satellite, facility footprint), OCCRP (cross-border community-impact reporting)	Verifies specific permit and emission numbers from primary documents rather than from outlet framing
3, taxpayers	State tax abatements, state capture, corporate lobbying	Transparency International (CPI + state-capture frameworks), ICIJ (elite-capture investigations), Integrity Index (Congressional lobbying conflicts)	Good Jobs First is the specialty US source; the canon adds the corruption-and-elite-capture framing and the federal-lobbying conflict-of-interest layer
4, investors	Hyperscaler capex vs revenue, bond market exposure	IMF (Global Financial Stability Report on AI-infra debt concentration), FiveThirtyEight data	Goldman is a US private-sector input, not multilateral; IMF GFSR is the site's preferred independent angle on systemic exposure
5, other businesses (input side)	TSMC / CoWoS allocation, GPU supply, grid capacity	Nikkei Asia (Asia-press lens on TSMC + Asian supply chains, per the independent-press-by-language rule), Bellingcat (open-source supply-chain analysis)	Asia-language press is the site's preferred lens on TSMC / GPU production and CHIPS Act packaging facilities; US tech-press citations are useful but ownership-tilted
5, other businesses (demand side)	Office-vacancy cascade, CRE-loan stress, municipal tax-base erosion, supporting-business failure, white-collar labor displacement	IMF (GFSR on CRE-loan systemic risk + WEO on labor displacement), FiveThirtyEight data (WFH / occupancy datasets), Reuters / BBC / Guardian / Economist (global wires on CRE), Nikkei Asia (Tokyo / Singapore CBD parallel)	Bank-balance-sheet framing is incomplete; the actually-exposed surface includes small-business owners, municipal bondholders, and the displaced workers themselves. IMF GFSR is the site's preferred independent angle on CRE-loan concentration risk
Ouroboros / labor-capital velocity	Private-equity simultaneous CRE + AI exposure; MPC asymmetry between labor wages and capital returns	IMF (WEO on inequality and aggregate demand), ICIJ (offshore finance + capital-acquisition structures), Transparency International (tax-preference state-capture), Integrity Index (Congressional lobbying on capital-gains and CRE policy)	The velocity-asymmetry argument is canonical macro (MPC; Piketty / Atkinson / Saez on wealth concentration); IMF WEO + GFSR cover the multilateral macro angle. The PE ouroboros itself is not yet in canon coverage
Implicit backstop / moral hazard	2008 TARP + AIG, 2019 repo-market intervention + Standing Repo Facility formalization, 2020 CARES + Fed corporate credit facilities, 2023 SVB systemic risk exception, ongoing CRE forbearance and quiet regional-bank rescues	IMF (GFSR macroprudential / systemic-risk analysis), Transparency International (state-capture via tax and regulatory preferences), Integrity Index (Congressional lobbying on financial regulation + Dodd-Frank carve-outs), FiveThirtyEight data (federal-spending and bailout-outlay datasets)	The bailout pattern is documented in primary Treasury / FDIC / Fed releases; the canon's macroprudential angle is IMF GFSR. The "exception-based, not rule-based" character of post-Dodd-Frank rescues is the structural read
Commercial state (vendor + customer ROI)	NVIDIA / OpenAI / Anthropic profit + loss; enterprise AI productivity by function	IMF (WEO on AI's productivity contribution to GDP), Reuters / BBC / Guardian / Economist (global wire on vendor earnings), Nikkei Asia (TSMC / SK Hynix / Samsung supply-chain margin reality on NVIDIA's profit pool)	NVIDIA's filings are SEC-public. OpenAI / Anthropic are private and depend on leaked filings + reporting; McKinsey / BCG / Gartner survey data is private research, not canon. The canon does not directly aggregate enterprise-AI ROI, this is a gap
Hardware moat erosion (local inference)	DGX Spark, Strix Halo, RTX 50-series, RTX PRO Blackwell; open-weight model capability vs cloud frontier	Nikkei Asia (TSMC / SK Hynix / Samsung supply-chain coverage on consumer-AI hardware), Bellingcat (OSINT verification on hardware-shipment claims), FiveThirtyEight data (open datasets)	NVIDIA's product announcements (CES 2025, GTC 2025) and SEC filings are primary. Hardware-capability benchmarks live in the technical press (Hugging Face, llama.cpp project, MLPerf) which are outside canon; cited inline
All siphons	Viral / contested claim verification	Snopes, Ground News (bias + coverage check)	Quick gut-check before any number gets treated as fact; Ground News surfaces how a story splits across the political spectrum

Where canon coverage runs thin (and what filled in):

• State-level tax-abatement accounting: the canonical aggregator is Good Jobs First; no canon equivalent. Their methodology (per-state GAAP disclosure or proxy from announced subsidy values) is the only systematic source.
• Facility-level NOx permit applications: Mississippi Free Press, Virginia Mercury, Sierra Club. Bellingcat could verify any individual claim via permit-document OSINT but does not maintain a running aggregate.
• EchoLeak-class commercial security incidents: canon doesn't have a security-disclosure aggregator; relies on Aim Security / vendor advisories / CVE feeds.
• Major-city CBD office vacancy and CRE-loan stress: the canonical aggregators are CBRE, Cushman & Wakefield, and JLL (private commercial-real-estate firms). No canon equivalent; IMF GFSR covers the macro systemic-risk angle but not the per-market vacancy detail.
• Private-equity simultaneous CRE + AI exposure (the ouroboros): the PE GP-side disclosures are sparse and the LP-side reporting is private. The canon's investigative outlets (ICIJ, OCCRP) have covered adjacent PE state-capture but not the specific CRE-vs-AI internal-portfolio contradiction. The argument as presented is a structural read assembled from public AUM disclosures and academic macro on MPC asymmetry.
• Enterprise AI ROI by function: the canonical aggregators are McKinsey Global AI Survey, BCG GenAI Productivity Index, and Gartner's agentic-AI tracking, none of which are in the canon (all are private research firms). IMF WEO addresses AI's macro productivity contribution but not the per-function decomposition.
• Local-inference hardware capability benchmarks: the canonical aggregators are Hugging Face leaderboards, MLPerf, llama.cpp project benchmarks, and Apple-silicon LLM benchmarking communities (LocalLLaMA on Reddit, the "ggml" project). None of these are in the canon. NVIDIA / AMD product announcements and SEC filings are primary; the canon does not yet cover this segment.

These gaps are appropriately acknowledged in the per-siphon citations below.

Sources

CBRE, Cushman & Wakefield, JLL Q1 2026 US office market reports (CBD vacancy by major market). IMF Global Financial Stability Report (CRE-loan systemic risk; bank Tier-1 concentration). FDIC bank failure / CRE concentration tracking. San Francisco Office of the Controller, Joint Report on the Five-Year Financial Plan (FY2025-26 through FY2029-30), with the November 2025 update placing the projected two-year general-fund deficit at roughly $937M, driven primarily by commercial-property reassessments (corroborated by SF Standard, November 2025; Real Deal, January 2026). NYCB 2024 CRE-driven loss-loan reserve disclosures. Goldman Sachs, "Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out." Fortune, "Middle-class Americans paying for the data center and AI boom" (Goldman Sachs, Feb 2026). CNBC, "Who is really footing the AI energy bill?" (Mar 2026). Good Jobs First, "Data Center Tax Breaks Becoming Billion-Dollar Budget Sinkholes" and "Cloudy with a Loss of Spending Control." The Register, "US states can't account for datacenter tax breaks. Literally" (Apr 2026). Tom's Hardware, "Big Tech AI spending plans reach $725 billion." CNBC, "Big Tech capex topping $1 trillion in 2027." Fortune, "Tech may only get half the profit needed to justify AI investment." Breckinridge Capital Advisors, "The Price of AI." Tech Insider, "U.S. AI Data Center Delays: 7 GW Capacity Crisis." World Economic Forum, "Is power grid connectivity the strategic bottleneck for AI?" Fortune, "America's data centers are thirsty. Rural towns are paying the price." Lincoln Institute, "Data Drain: The Land and Water Impacts of the AI Boom." AEP Ohio, "Data Center Tariff Update." Newsweek, "How Data Centers Are Set To Impact Home Value." Priest (SSRN), "Not In My Back Yard: Effects of Data Centers on Housing Prices." American Farm Bureau Federation, "Balancing Data Center Growth with American Agriculture." Inside Climate News, "Data Centers' Use of Diesel Generators." Sierra Club, "Looser Rules, Dirtier Air." Mississippi Free Press, "Amazon's Canton AI Data Center Brings Dust, Noise, Pollution Fears." EESI, "Data Centers Are Contributing to PFAS Forever Chemical Pollution." Virginia Mercury, "What's in the water?" US News, "Living in Hell: Virginia Residents Grapple with Data Center Noise, Air Pollution." Frontiers in Climate, "Health implications of the rapid rise of data centers in Virginia." Spheron, "GPU Shortage 2026." Kavout, "Why are NVIDIA H100 GPU rental prices surging 40%." VentureBeat, "Why enterprise GPU utilization is stuck at 5%." Federal Reserve Board, "Statement Regarding Repurchase Agreement Arrangements" (28 July 2021, establishing the $500B-aggregate SRF) and the New York Fed's Standing Repo Facility FAQs for ongoing counterparty expansion. Bloomberg, "Fidelity to Cut 800 Staffers as It Overhauls Tech, Product Teams" (May 2026); Fortune, "Amazon CEO Andy Jassy says he's cutting middle managers..." (March 2025); Entrepreneur, "Amazon CEO Andy Jassy Says He Wants Fewer Middle Managers" (October 2024); Microsoft Q3 FY2025 earnings call transcript (April 2025, Amy Hood "reducing layers with fewer managers"); CNBC, "Meta announces 5% cuts in preparation for 'intense year'" (January 2025); The Next Web / CNBC, "Meta cuts 8,000 jobs amid record Q1 revenue as Zuckerberg bets $145 billion on AI infrastructure" (May 2026); Fortune, "Salesforce CEO Marc Benioff says his company has cut 4,000 customer service jobs as AI steps in" (September 2025); Digital Applied / Solutions Review on Klarna's 2026 AI-layoff reversal; White House, Executive Order 14318: Accelerating Federal Permitting of Data Center Infrastructure (23 July 2025); DOE press release on Idaho National Laboratory / Oak Ridge Reservation / Paducah Gaseous Diffusion Plant / Savannah River Site selections (24 July 2025); White House Stargate joint-venture announcement (21 January 2025). Institute on Taxation and Economic Policy, "Four Big Tech Companies Avoid $51 Billion in Taxes in Wake of One Big Beautiful Bill Act" (2026); ITEP, "From 0% to 1.2%: Amazon Lauds Its Minuscule Effective Federal Income Tax Rate" (2022); ITEP, "Corporate Tax Avoidance in the First Five Years of the Trump Tax Law"; Common Dreams, "Trump-GOP Law Slashes Amazon's Tax Bill by 87% as Company Fires 30,000 Workers, Profits Soar" (2026); Microsoft Corp. Form 8-K (October 2023) disclosing $28.9 billion IRS NOPA for tax years 2004 to 2013 (Puerto Rico cost-sharing transfer-pricing dispute); Bloomberg Tax, "Microsoft's $29 Billion Tax Bill Offers Transfer Pricing Lessons"; Congressional Research Service R47328, "The 15% Corporate Alternative Minimum Tax" (CAMT background and JCT scope estimate of ~150 affected companies); One Big Beautiful Bill Act (Public Law 119-XX, July 2025) for the immediate-expensing and bonus-depreciation provisions applied to AI-infrastructure capex. Cresa Q1 2026 Washington DC Office Market Report; Avison Young Downtown Chicago Office Market Report; CBRE Q1 2026 US Office Market Report; CommercialCafe National Office Report, May 2026 (US national, LA, Bay Area, Houston downtown vacancy). NVIDIA Q1 FY2027 Financial Results (NVIDIA Investor Relations, May 2026).