SambaNova Marks Up to $11B as AI Chip Rounds Reflate 4-8x

New York built the mechanism that makes this contagious: an executive order blocking environmental permits for 50MW+ facilities, and a legislative bill at 20MW+ waiting on Hochul's signature. The politics run 46% for, 21% against, on utility bills and grid stress that exist in every populous state, so twelve-plus states already have comparable bills filed. When the binding constraint stops being chips (buyable) and becomes permits and power contracts (not accelerable at any price), value migrates to whoever is holding pre-permitted capacity, behind-the-meter gas, grid interconnect. One hyperscaler campus needs seven new natural-gas plants. That is the whole sentence.

Demand delivered its own verdict. IBM warned Q2 would disappoint because clients diverted mainframe budgets to AI hardware, and the stock fell 23% in a single session — its worst drop in decades — to $204B, which is below Palo Alto Networks and below CrowdStrike. The same reallocation that rewards the winners now punishes the perceived losers on the same afternoon. Anything AI-adjacent but not AI-native — legacy IT, generic enterprise resellers, consulting-flavored transformation plays — is a re-rating candidate on its next print, or rather on the first print the market reads carefully.

The memory squeeze completes the picture. DRAM and HBM reallocated to data centers drove smartphone shipments down 11% to a 13-year low, and only Samsung at 24% and Apple at 20% gained share while the mid-tier got euthanized. That is demand destruction, which is the tell that the reallocation is physical rather than narrative. AWS is constrained enough that startups are defecting to alternative clouds, which opens a time-boxed logo-grab window for neoclouds and inference specialists. Time-boxed being the operative word.

The buildout continues in aggregate — China AI exports up 27%, Intel committing €5B to an Irish fab — so the argument is not whether but where value accrues, and the sources disagree violently. The resolution is a barbell: scarce physical inputs on one end, capex-light AI monetizers on the other, capex-heavy middlemen and legacy IT squeezed in between. This is probably wrong on the timing. It is likely right on the shape.

The coordination is the new part, and the data underneath it. Nadella, Benioff, and Karp all warned inside 72 hours that Anthropic and OpenAI harvest customer knowledge to compete up the stack — Karp put it bluntly: they'll 'take the alpha of my business, transfer it into their model weights, and compete against me.' The messaging is self-serving, of course (Microsoft has no frontier model; Salesforce runs on both providers), but the anxiety it channels is a real demand signal for a neutral middleware layer. Dedicated instances are already 'the norm' at financial institutions. That entrenches the hyperscalers as the toll booth.

The margin data hardened this week. Open-weight models are 29% of Vercel Gateway token volume at under 4% of spend — a 7-to-1 volume-to-revenue gap, which is the floor collapsing under commodity inference. Meta enters the API market on price with a 1M-context model, Zuckerberg calling current models 'too expensive.' Distilling a frontier model into a deployable specialist runs about ~$250. Three labs shipped computer-use in the same week; three coding agents shipped identical /goal features inside a fortnight. Feature moats now have a half-life under six months.

Value sticks at distribution and evals, or rather, wherever the customer can't easily switch. Codex hit 6M users adding 1M/day, roughly 3x Claude Code's last-reported figure, which says the coding-agent war is won on distribution and not on benchmarks. a16z planted its flag: pure SaaS is 'nearly uninvestable,' and the next trillion accrues to transformation companies that encode incumbent processes into agents, with proprietary eval suites as the moat (coding is ~99% of AI revenue because code self-evaluates). Read the memo as a preview of their next checks. Read it also as a leading indicator of multiple compression for eval-less subscription software.

The diligence change is smaller and more useful: token-level COGS is now a required IC input. Identical workloads cost 73% more on one vendor's tokenizer than another's, and agent loops multiply consumption on top of that. Headline ARR on a thin wrapper can sit over structurally negative unit economics. Read the footnote.