Valar's $5B Mark Tests Nuclear-for-AI vs Oklo's 40% Drop

The Helios spec sheet is not the story, since full-stack launches are cheap and everyone has one. The customer list is the story. Microsoft Azure committed to deploying Helios for frontier-model inference in H2 2026, wrapped in an actual product (Azure Foundry Managed Compute), and it did so alongside Meta, OpenAI, and Oracle, which is to say the four deepest-pocketed and most Nvidia-dependent buyers in the market are hedging in public and paying for the privilege. Helios runs roughly $5–5.5M/unit against $3.5–4M for Nvidia's Vera Rubin, and AMD holds about 4.5% share against Nvidia's more than 95%. That is not price parity so much as a diversification premium, and one the buyers appear happy to pay.

Three more vectors point the same way, with different clocks. Etched has pulled 400+ engineers from Nvidia and TSMC and carries $1B in demand before shipping, the kind of talent flight plus pre-shipment pull that sometimes precedes an infrastructure re-rate and sometimes precedes nothing. Google's Frozen v2 targets 6–10x tokens per watt by hardcoding Gemini into silicon, though not until 2028, which is a long time in this market. And China is compounding faster than assumed: Zhipu/Z.ai has stood up a 1GW all-domestic-chip data center now, which undercuts the 18–24 month China lag baked into most 2024-era decks.

Where the sources agree: value is migrating from merchant silicon toward whoever owns the architecture, the network, or the distribution. Where they hedge, sensibly: 2028 timelines slip, and CUDA lock-in has outlasted hardware challengers before. Nothing here breaks the moat this year, and it would be a mistake to trade as if it did. What it breaks is the premise. The 'there is no alternative' argument underwriting Nvidia's pricing power, and the elevated multiples across its supply chain, is now empirically contestable, which is a different thing from wrong.

The nearer-term trade sits a layer below the hardware headlines, in cost-aware routing and orchestration middleware. Ramp's internal router cut inference costs 30% with no performance loss, and enterprises keep blowing their AI budgets even as per-token prices collapse. That is demand-side evidence for a layer that is not the model at all.

Disclosure discipline is the story in OpenAI's confidential IPO filing, which puts the company alongside Anthropic and SpaceX in an explicit race to public markets and, more usefully, forces the first real margin transparency onto frontier AI. That matters more than whatever valuation gets attached, because leaked figures already show OpenAI losing billions annually despite revenue growth, and ChatGPT's share below 50% as Gemini and Claude take theirs. The inevitable-dominance narrative is repricing.

Capital, in the meantime, is sorting along lines that do not resemble each other. Foundation-model liquidity is the IPO trio's game. Physical-AI mega-capital is a different bet entirely: Bezos-backed Prometheus raised $12B pre-revenue for an 'artificial general engineer,' which resets the venture-scale bar and means undifferentiated robotics plays are now competing against a war chest rather than against each other. Then there is China compute self-sufficiency, where DeepSeek took $7B in its first external round on the strength of claims that a Huawei team post-trained a 1.6-trillion-parameter model on 1,000 domestic Ascend 910C chips. If that claim holds, the export-control moat under several Western chip positions is thinner than their prices assume.

Underneath all of this sits an asymmetry worth pricing: Google pays SpaceX $920M/month for compute while cutting consumer AI subscription pricing, which means input costs stay high even as prices fall, a fact that belongs in the unit economics of every AI subscription business and mostly is not there yet. Separately, reported US-government interest in equity stakes in AI giants, plus Anthropic's dual role as commercial lab and regulatory advocate, means control terms on any new frontier-adjacent deal now carry a governance-risk premium that was not in last cycle's term sheets.

The edge, such as it is, is sequencing: public comps become real this cycle, and positioning ahead of each split beats reacting after the S-1s force the market to do the math. There is a version of events in which the filings stall or the leaked losses shrink faster than anyone expects and the dominance narrative reprices back upward, and that version is not impossible. It is just not what the disclosures suggest.