Arming both sides

In a gold rush, don't dig for gold. Sell shovels.

Here is a number that should reframe how you read every AI headline. In calendar Q1 2026, NVIDIA earned $58.3 billion in net income in a single quarter, on $81.6B revenue, with data-center revenue of $75.2B, up 92% year over year. Hold that quarter flat and the annualized run-rate implies roughly $230B in net income from one company. That is a 4x extrapolation, not guidance. The $58.3B quarter is reported fact.

Now put it next to the people buying those chips.

• OpenAI: about $2B per month in revenue (~$24B annualized), and a projected $14B loss in 2026, $27B cash burn, with its own filings telling investors to expect material losses through 2028 and no profit until 2030.
• Anthropic: the bright spot. It crossed roughly $47B annualized by late May 2026, up from $9B at year-end 2025, with burn forecast to fall toward 9% of revenue by 2027. The only model lab on a credible near-term path to positive cash flow.
• The Big Four (Microsoft, Google, Amazon, Meta): committing ~$725B in 2026 capex, a 77% jump over 2025, against roughly $450B in consensus revenue for the same year. The capex is about 60% larger than the revenue it is supposed to justify.

So the profit pool of the entire AI boom is, right now, concentrated almost entirely in the one firm that does not run a model, does not rent a token, and does not own a data center. Everyone downstream is either burning capital or betting that revenue shows up later.

This is the oldest pattern in boom economics. In the 1849 gold rush most of the miners went home broke; the durable fortunes were made by the people selling them the picks, the shovels, and the blue jeans. Samuel Brannan cornered the supply of shovels around San Francisco and resold them at a brutal markup, becoming California's first millionaire without ever working a claim. Levi Strauss sold the denim. The lesson hardened into an investing maxim: in a gold rush, you want to own the supplier, not the speculative end of the dig. NVIDIA is the shovel-seller of the AI rush, and that much is widely understood.

That is the part most people already half-know. Here is the part they do not. Brannan only ever sold shovels. NVIDIA is also selling the machine that lets you skip the river.

NVIDIA is building the thing that breaks its customers' model, on purpose

The cloud-AI thesis is, structurally, a hardware-rental business. You pay per-million-tokens because owning the hardware to run a frontier model locally was infeasible. The entire $725B capex bet assumes that stays true.

At GTC Taipei on June 1, 2026, NVIDIA unveiled the RTX Spark superchip, the N1X SoC co-designed with MediaTek: a 20-core ARM CPU paired with a Blackwell GPU, up to 128GB unified memory, roughly 1 petaflop of FP4, on TSMC 3nm. It ships this fall in 30-plus Windows laptops and 10-plus desktops from Dell, HP, Lenovo, ASUS, MSI, and Microsoft Surface. NVIDIA's stated capability: run 120-billion-parameter models with 1M-token context, locally.

Twelve months ago that capability meant cloud rental at roughly $30K-equivalent monthly cost. This fall it is a consumer laptop SKU.

Memory bandwidth, not capacity, is the ceiling

This is the part that decides how much of the cloud's business actually walks out the door, so it is worth slowing down for. The erosion is real but it is selective, and the reason is a single piece of hardware physics.

Start with what "running a model" costs the machine. To produce each word (more precisely each token, roughly three-quarters of a word), the computer reads the model's weights, the billions of numbers it learned during training, out of memory and does math on them. Do that fast enough and text streams out at reading speed. The rough line where local AI stops feeling sluggish and starts feeling like a cloud service is about 30 tokens per second, a little faster than you can read along.

Two different specs get conflated here, and the gap between them is the whole story:

• Memory capacity is how big a model you can hold at once, in gigabytes. This is the number the marketing leads with: "128GB unified memory."
• Memory bandwidth is how fast you can read those weights back out, in gigabytes per second. This is the number that actually sets your words-per-second.

Think of capacity as the size of a warehouse and bandwidth as the speed of the conveyor belt pulling stock out of it. A bigger warehouse lets you store a bigger model. It does nothing for how fast you can run one. And on these consumer machines the belt is slow: the unified-memory tier (RTX Spark, DGX Spark, AMD's Strix Halo) reads at roughly 250 to 300 GB/s. A datacenter H100 reads at about 3,350 GB/s. That is a 5x to 7x gap, and for the wrong kind of model it maps almost directly onto how fast words come out.

A dense model is the wrong kind. "Dense" means every parameter takes part in every token, so the machine has to haul the entire weight set off the belt for each word it generates. Run a dense 70-billion-parameter model on one of these laptops and you get roughly 5 to 10 tokens per second: readable, but well under the 30 threshold and painful for real work. So the headline that you can fit a 70B model on a laptop is true and misleading at the same time. You can fit it. You just can't run it fast. Dense 70B and up still belongs in the cloud, where the belt moves 7x quicker.

Mixture-of-Experts (MoE) models are built to dodge exactly this bottleneck. Instead of one monolithic network where every parameter fires on every token, the weights are split into many smaller specialized sub-networks (the "experts"), and a lightweight router sends each token to only a few of them. The full model still has to be stored, so you still need the capacity, but only a small slice of its weights gets read for any given token, so you barely tax the slow belt. Qwen3-Coder 30B-A3B carries 30B parameters in total but activates only about 3B per token, so its per-token bandwidth cost is roughly a tenth of a dense 30B model's. On the same Strix Halo box that crawls through a dense 70B at single digits, it posts 70 to 100 tokens per second. Gemma 4, DeepSeek V4 Flash, and Llama 4 Scout work the same way. NVIDIA's "120-billion-parameter model on a laptop" claim is true because the target models are MoE: a dense 120B would be unusably slow on a 300 GB/s machine, but a sparse one that only reads a few billion weights per token is not. And the labs are publishing MoE-default open weights faster every quarter, which is precisely the slice of demand that can leave the cloud.

The software side stopped being a barrier a while ago, and the switching cost is close to zero. The standard way an application talks to OpenAI is an HTTP request to its API. Tools like Ollama (free, open-source) expose that same interface from a server running on your own machine, so you change one URL, from OpenAI's endpoint to http://localhost:11434, and the same application stops paying per token and starts running locally for free, with nothing else rewritten. LM Studio packages the identical capability as a desktop app for non-developers. llama.cpp provides the low-level engine for every consumer chip family, and the popular AI coding tools (Cursor, Continue, Aider, Zed) already support local backends. The only thing the cloud was ever bottlenecking was the silicon. The silicon showed up.

Code generation is the test case, and the moat is thin

The clearest place to watch this play out is AI coding assistants. Writing code is the highest-revenue use of AI today and the most heavily benchmarked, so it is where the cloud's quality advantage should be most defensible. It is not. Open models built for code (Qwen3-Coder, DeepSeek-Coder V3, GLM-4.6) now match GPT-4-class quality at sizes that fit on the local hardware described above.

Here is the part worth sitting with if you hold the stocks. The model was never the hard part. What makes an AI coding tool genuinely useful is the software wrapped around the model: the scaffolding that lets it edit files, run your tests, pull the relevant slice of a large codebase into view, and feed its own error messages back in to try again. A paid product like Cursor, or a cloud API subscription, is largely selling that scaffolding bundled with the model. Every layer of it now exists as free, open-source software that runs locally. A developer pairing the open tool Aider with a local Qwen3-Coder model on a $4K desktop has, for everyday in-house coding work, what required GPT-4 plus Cursor plus per-token cloud billing eighteen months ago. The thing customers were paying a subscription and a metered per-token fee for is increasingly a thing they can run themselves for the one-time price of the hardware.

Only one side of this trade is hedged

This is a hedging story, and the asymmetry is the whole point. NVIDIA is hedged across every layer: cloud (H100/B200/GB200), data-center workstations (DGX), prosumer (DGX Spark), consumer desktop (RTX 50-series), and now laptop chips. The hyperscalers are not. Their $725B was placed on the premise that enterprises and consumers cannot run useful AI locally and must rent it. As that premise erodes for the MoE-served bulk of inference demand, the cloud's addressable market shrinks to the cases where renting still pencils: bursty load, dense-frontier workloads, regulated compliance. That is a much smaller market than universal cloud lock-in.

Add the depreciation knife Goldman keeps flagging. Shorten modeled chip life from 5 years to 3, closer to real hyperscaler utilization, and you add about $1 trillion to annual industry depreciation, enough to absorb most of the optimistic-case revenue growth the capex depends on. Goldman's own analysts warn hyperscalers may realize only about half the profit needed to justify the spend.

NVIDIA wins whichever way it breaks

NVIDIA's interests align with whichever side wins. The hyperscalers' interests align with only one outcome. NVIDIA sells the H100s the capex bought and the laptop chips that obsolete the per-token billing model those H100s were bought to enforce. It collects if the cloud lock-in holds, and it collects if the cloud lock-in breaks.

Brannan sold shovels to whoever was digging, and it did not matter to him whether any given miner struck the vein or went home empty. NVIDIA has gone one better: it sells the shovels and the machine that drains the river, and it is paid on both. When one company is the only guaranteed winner of a trillion-dollar bet, that is not a vote of confidence in the bet. That is the house telling you who actually owns the table.