Market and trends

5 min read · Chapter 6 of 10

Who holds the power

Compute is the binding constraint, and power follows whoever controls a piece of it.

Clouds and the grid. Alphabet said in April 2026 that it was "compute constrained in the near-term", Microsoft's CFO said in July that it "remains capacity constrained", and AWS called its Trainium2 chips "largely sold out". The binding input is moving from chips to power: CoreWeave had about 1.5 GW active against 3.7 GW contracted in June 2026. PJM, the grid operator for much of the eastern US, cleared its capacity auction at the price cap three times running and ended about 6,800 MW short of its reliability target in July 2026, with data centers the main driver of new load.
Nvidia allocates the scarcest chips, keeps the highest margin and invests in labs, GPU clouds and providers. TPUs, Trainium, AMD and Broadcom's custom chips check it in part.
Labs over builders. Labs set prices, limits, retention and retirement dates, and ration when capacity is short: peak-hour cuts in Claude Code in March 2026, Priority Tier withdrawn, OpenAI limiting GPT-4.5 to its top plan in 2025 because it was "out of GPUs". Downstream, GitHub paused Copilot Business signups in April 2026 for lack of capacity. When capacity lands, limits rise, as Anthropic's did on May 6, 2026.
Clouds and labs. The clouds are the labs' suppliers, investors, distributors and competitors, and the labs are the anchor tenants behind the clouds' backlogs.
Contracts. Anthropic's commitments of at least $518 billion run about a decade, around 80% non-cancelable, per the leaked draft filing as reported. OpenAI's are reported at about $600 billion through 2030, including more than $300 billion at Oracle and a 750 MW deal that Cerebras's own filing values at more than $20 billion. Meta signed about $21 billion more with CoreWeave in April 2026. xAI builds its own Colossus sites, which is why Anthropic could rent capacity from SpaceX on 90 days' notice.
Suppliers as investors. OpenAI's $122 billion round at $852 billion, closed around April 2026, included $30 billion from Nvidia and $15 billion from Amazon (plus $35 billion conditional) tied to 2 GW of Trainium. Amazon marked its Anthropic notes up from $42.2 billion to $97.9 billion in one quarter, according to an analysis of the leaked filing. These loops prop up demand signals and concentrate risk.
Providers and gateways compete on speed, price and switching, with thin margins; model makers can shame hosts (Moonshot found gaps between third-party and official Kimi APIs "widespread"), and gateways are being folded into security companies.
Builders hold demand and little contractual power. Their leverage is spend commitments, a second provider and moving easy traffic to cheaper models.

Tailwinds and headwinds

These are the forces I'd expect everyone in this industry to watch over the next couple of years, and each one helps some players and hurts others. Several are the same scarcity seen from different sides: capacity arriving is a tailwind, and the power and debt behind it are headwinds.

Tailwinds:

What's changing	Who gains	Who loses
Each task uses far more tokens as agents, reasoning and coding grow, and in Ramp's spending data the average business's monthly token spend rose 13 times from January 2025.	Labs with the strongest agentic and coding models (Anthropic, OpenAI, Google DeepMind); the clouds that resell them; Nvidia, CoreWeave and Oracle; gateways, as budgets and caching become necessary	App companies on flat seat prices, whose costs rise with use; builders without caching or budgets
New data centers and chips arrive through 2026 and 2027, so revenue the clouds say is capped by supply can grow.	Whoever energizes capacity first (Microsoft, AWS, Google, Oracle, CoreWeave, Nebius, Crusoe); labs that have been rationing (Anthropic, OpenAI); builders, whose rate limits rise	Probably inference providers whose edge was spare capacity (Fireworks, Together AI, Baseten); sellers of spot GPU time
Big companies are moving from pilots to production, and 37% of US firms with 250 or more employees now use AI somewhere in the business.	Labs with enterprise sales and indemnities (Anthropic, OpenAI, Google DeepMind); the cloud platforms that carry the contract; gateways sold as governance (Portkey, Cloudflare AI Gateway)	Labs without an enterprise channel or a Code of Practice signature (DeepSeek, Moonshot, Alibaba's Qwen) with regulated buyers; open weights' share of enterprise dollars
The three big clouds now sell most labs' models, so model spend draws down the buyer's cloud commitment.	AWS, Microsoft, Google; labs sold through several clouds; enterprise buyers, who get one contract	Labs' direct-sales margin; inference providers and gateways competing with the clouds' own catalogs and routers (Together AI, Fireworks, OpenRouter); labs missing from a buyer's committed cloud
Google's TPUs, Amazon's Trainium, AMD and specialist chips give the labs more supply beyond Nvidia.	Google, Amazon, Broadcom, AMD, Cerebras; labs that run on several kinds of chip (Anthropic, OpenAI)	Nvidia's share of new spending, though not yet its revenue; probably GPU clouds that run only Nvidia (CoreWeave, Nebius)

Headwinds:

What's changing	Who gains	Who loses
Power and grid connections now limit how fast capacity grows more than chips do, and PJM's July 2026 auction ended about 6,800 MW short of its reliability target.	Whoever already holds powered sites and grid connections (Microsoft, AWS, Google, Oracle, xAI's Colossus); developers that bring their own power, such as Crusoe	GPU clouds with contracted but unbuilt capacity (CoreWeave, Nebius); every lab whose capacity arrives late; builders, through rate limits
Labs and GPU clouds have signed decade-long, mostly non-cancelable compute contracts and loans that only work if revenue keeps climbing.	Chip makers already paid (Nvidia); clouds with many customers (AWS, Microsoft, Google); builders, if a downturn floods the market with capacity	GPU clouds with debt and a few big customers (CoreWeave, Nebius, Crusoe); the labs with the largest commitments (Anthropic, OpenAI); investors in deals where suppliers fund their own customers
Labs are protecting margins with new top tiers, scheduled price steps and pricier old models, so builders stop getting cheaper tokens at the top.	Labs' margins (Google, OpenAI, Anthropic, Zhipu); open-weight hosts as the cheaper exit (Fireworks, Together AI); gateways that make switching easy (OpenRouter, LiteLLM)	App companies on fixed prices; builders tied to one model; heavy users of retired models kept on at the clouds
Renting GPUs and buying memory got dearer in 2026, with one-year H100 contracts up about 40% in six months.	Owners of installed chips and capacity on long contracts (Nvidia, AMD, the hyperscalers, GPU clouds)	Inference providers on thin margins and short contracts (Together AI, Fireworks, Baseten, Cerebras); self-hosters
Cheap open-weight models keep prices low for everyday tasks, at a half to a sixth of a closed model's price.	Builders; inference providers that win on volume (Fireworks, Together AI, Baseten, Groq, Cerebras); open-weight makers' reach	Closed labs' small and mid tiers (OpenAI, Anthropic, Google, Mistral, xAI); providers themselves, on price per token

The one thing I'd watch is the gap between contracted and powered capacity in the GPU clouds' quarterly reports, because it shows first whether rate limits will ease and whether the long contracts can be paid.