The Watts War

On May 6, Anthropic announced an agreement with SpaceX for access to the entire compute capacity of Colossus 1, the Memphis supercomputer that xAI built and SpaceX inherited when it absorbed xAI in a February stock-for-stock combination at a $1.25T combined enterprise value. The numbers are over 220,000 NVIDIA GPUs (a mix of H100, H200, and GB200 per prior xAI disclosures), drawing more than 300 megawatts of new capacity, online within the month. As a downstream consequence, Claude Code’s five-hour rate limits doubled for Pro, Max, Team, and seat-based Enterprise; peak-hour throttling on Pro and Max disappeared; and Opus models saw API rate limits raised. If you write code with Claude all day, that is the part you actually feel. It is also not the story.

The story is that ten weeks ago Elon Musk was calling Anthropic “Misanthropic”, accusing them of “stealing training data at massive scale,” and tweeting that the company “hates Western Civilization”. In March he was mocking Dario for talking about AI consciousness. In May he is selling them watts. That is the entire industry compressed into one transaction. The public bile evaporated against the gravity of a Memphis power substation, and it evaporated fast. The commercial pull was stronger than the political performance.

To see why a Memphis substation outranks a feud, look at what April actually shipped. Opus 4.7 landed April 16, posting 87.6% on SWE-bench Verified at $5 in / $25 out per million tokens with a 1M-token context. The GPT-5.5 family came out a week later and retook the top of the board at 88.7%. DeepSeek V4 followed the next day, a 1.6-trillion-parameter MoE running roughly seven times cheaper than Western frontier ($3.48 per million output tokens against Anthropic’s $25), trained on Huawei silicon. Qwen 3.6 from Alibaba now leads the open-weights segment at 27B dense and 35B-MoE variants, the size band where most production deployments actually live. Two weeks earlier, Meta had shipped Muse Spark, its first closed-weights flagship out of Meta Superintelligence Labs, breaking with the Llama open-source lineage. Three frontier-tier coding releases inside an eight-day window, plus an open-weights wave at the deployment size, plus a closed-weights flag-planting from the historically open incumbent.

When five labs are clustered that tightly on capability, capability is not the binding constraint anymore. Capability has commoditized. The places where the labs differ now are upstream of the model weights and downstream of the API: power contracts, foundry slots, legal cover, and which sovereign accepts your training corpus. That reframing changes how you read every announcement. The Colossus deal is not a model announcement. It is a watts announcement, and watts are what is scarce.

Run the inventory. Anthropic now holds up to 5 GW with Amazon (Trainium-heavy Project Rainier), 3.5 GW of Google TPU through Broadcom starting 2027 on top of an earlier Google Cloud TPU commitment, the $30B Microsoft / NVIDIA Azure reservation, $50B with Fluidstack, and now Colossus. That is roughly 10 gigawatts of contracted compute spread across four substrates and four counterparties, which is portfolio construction, not a procurement strategy. It is also a hedge against any single chip vendor or hyperscaler getting a chokehold on the supply. The Colossus piece in particular is interesting because it gives Anthropic a slug of NVIDIA H200/GB200 capacity that is already physically deployed and pulling power, which the AWS and Google buildouts will not match for quarters yet. You rent Colossus when you cannot wait for your own concrete to cure.

Meanwhile, Cerebras filed an S-1 on April 17 targeting roughly $26B at the top of its $115–$125 range (its pre-IPO Series H pegged it at $23B), anchored by an OpenAI Master Reservation Agreement worth more than $20B and covering 750 MW expandable to 2 GW. That filing landed into a publicly identified gap of seven gigawatts between announced AI data-center buildout and actually-under-construction capacity for 2026. The market is being asked to fund expansion into a power deficit that the utilities themselves have flagged. PJM, the regional grid operator that serves Pennsylvania and twelve other states, has launched a hunt for fifteen gigawatts of new generation against roughly five under construction, and Pennsylvania GOP incumbents are increasingly exposed on data-center siting because residential ratepayers are noticing the load. Watts are not just a procurement problem. They are turning into a campaign issue, and once a thing becomes a campaign issue in Pennsylvania it is no longer something you negotiate quietly with a utility.

The other side of the board is hardening too. China blocked Meta’s $2B Manus acquisition on April 27, the first publicly confirmed cross-border AI deal killed under the foreign-investment security review regime. Reading that as a one-off would be a mistake; it is the formal beginning of bidirectional decoupling, where you can no longer assume that a Western lab can simply buy its way into Chinese AI capability or the reverse. The EU’s GPAI enforcement powers under the AI Act activate on August 2. The Bartz copyright settlement fairness hearing is set for May 14 and will set a real number on training-data exposure for U.S. labs. Each of these is a piece of legal cover that some labs will have and others will not, and that asymmetry will show up in who can deploy where.

There is one line in the Anthropic announcement that I want to flag and then leave alone. Anthropic “expressed interest in partnering with SpaceX to develop multiple gigawatts of orbital AI compute capacity.” It is not contracted or even on a roadmap, just a press-release sentence that pairs the largest U.S. AI lab with the only operator whose launch cadence makes the claim non-absurd. I would not write a check against it, but I would not assume it is purely vapor either. Solar plus radiative cooling in vacuum starts to pencil out somewhere north of a few hundred megawatts of demand if your ground power is gated. Worth a mention; not real yet.

The same day the Colossus deal hit, Bank of America strategist Savita Subramanian flagged the impending SpaceX and Anthropic IPOs as an “issuance deluge” that would force passive funds to free capital and “smack of late-stage machinations.” That is what analysts say when valuations have outrun their own reasoning and they want a hedge in print. Anthropic is reportedly weighing offers at over $900B for a $50B round under discussion, which would put it above OpenAI’s $852B post-March round, and may run one last private round before an October IPO. None of this is unreasonable on its own merits if you accept the premise that the labs are essentially infrastructure companies with thin model wrappers, which is increasingly what they look like from the cap-table side.

So here is the claim worth disagreeing with. The next six months in AI will be decided almost entirely by who locked in which watts, which fab slots, and which jurisdiction’s legal cover, and almost not at all by which lab posts the highest SWE-bench number in June. The model layer is good enough across at least five providers that further capability gains have lost their commercial leverage; what differentiates a winner is whether they can serve the capability at price, at scale, in the geographies that pay. By that metric Anthropic just made a very large move, Cerebras is asking the public market to underwrite a power deficit, Meta has hedged with its first closed-weights flagship because openness costs more than it earns at the frontier, and the Chinese stack is now formally walled. The benchmarks will keep ticking up. That isn’t what the year turns on.