Sequoia, Thrive Back Nuclear for AI; Helion Marked at $15B

Two cost curves moved in opposite directions this cycle, and most three-year plans still treat them as one curve. On the model layer, Google shipped Gemini 3.6 Flash using 17% fewer output tokens at $1.50/$7.50 per million, a genuine deflationary push on agent inference. Underneath it, TSMC confirmed base price hikes of up to 10% plus a 10–15% HPC premium starting early 2027, which pushes some bills up roughly 25%. ASML raised prices 10%. Both cited demand described as 'very strong' for years, not quarters. Any cost model built on continuous deflation is now out of date at exactly the layer the deflation runs on.

The supply side is in open turmoil

A reasonable skeptic would note that the incumbent has outlasted challengers before, and the skeptic has history on their side. What the skeptic has to explain away is three fronts moving at once, each different in kind. Etched pulled 400+ engineers from NVIDIA and TSMC and holds $1B in pre-launch demand, which is a credible signal rather than vaporware. Taiwan authorities raided Supermicro and two supply-chain partners across nine sites and summoned six people in a widening probe, an active enforcement risk with real delivery-timeline consequences. And AMD's Helios rack system launched with Microsoft, Meta, OpenAI, and Oracle named. That is the first time hyperscalers have backed an NVIDIA alternative with production-scale intent.

The board-deck version of this story says inference is getting cheaper, so agent economics improve. The complete version is the netting exercise, and the netting is the actual work: model-layer savings are real, but they are being fought for on top of a hardware layer that is getting structurally more expensive. The tradeoff is now explicit. Single-vendor infrastructure buys simplicity at the price of concentration risk, and that price has moved from an IT decision to a board-level one. The vendor commitments made this quarter set the terms of the netting for the next two years.

AI is getting cheaper to run and more expensive to build. The gap between those two curves is where the next margin surprise is hiding.

The genuinely new development this cycle isn't the breach that's been circulating — it's what attackers are now aiming at. ENCFORGE ransomware was purpose-built to encrypt ~180 AI/ML model-weight extensions. Unlike transactional data, weights and checkpoints often have no clean restore path — encrypt them and you lose capability, not just records. Paired with JADEPUFFER's autonomous credential-theft-to-ransomware chaining, the threat surface is now the model itself.

This is a category flaw, not a vendor bug

Pillar Security showed Cursor, OpenAI Codex, Google's Gemini CLI, and Antigravity all share the same sandbox-escape pattern: a prompt injection hidden in a README or dependency causes the agent to write a file a trusted tool later executes outside the sandbox. That's not four unrelated bugs — it's evidence the current generation of coding agents shares an immature trust model, whichever vendor built it. Feature velocity has outrun security architecture across the category, which means any agent procurement made on capability benchmarks alone is incomplete.

The market is responding fast. Six credible defensive AI-security tools shipped in one cycle — from Anthropic, Datadog, Visa, Alpha-Omega, and evilsocket — and Google's orchestrated CodeMender pipeline found 55 V8 vulnerabilities versus 36 for Claude Opus 4.6, a signal that composite pipelines of specialized models beat single-frontier-model bets on real security work. Six overlapping entrants in one cycle means consolidation is near; evaluating now carries a leverage advantage that evaluating post-consolidation won't.

Model weights are now a ransomware target that can't be restored — if you lack a distinct backup tier, you're one unpatched CVE from losing capability, not just data.

None of this demands panic. It demands treating model weights and agent permissions as their own protected domain before an incident forces it.

SpaceX's reported $1.75T IPO was followed by a $60B acquisition of Cursor — if the reporting holds, not a coding-tools story but a deliberate vertical-integration move to arm xAI with talent, data, and distribution against OpenAI and Anthropic. Read alongside ChatGPT's share falling below 50% while OpenAI reportedly loses billions annually, the era of single-vendor AI dominance is ending — replaced by aggressive consolidation on one side and fragmentation on the other.

The commoditization underneath is the real accelerant. Alibaba-backed Kimi K3 (reported at 2.8T params, $3/$15 per M tokens) is claimed to beat Claude Fable 5 and GPT-5.6 Sol on the Frontend Arena benchmark — if the benchmark claim holds, evidence the capability gap between US and Chinese frontier labs is narrowing faster than most procurement cycles are built to notice. Switching costs that once locked enterprises into frontier providers are collapsing exactly as the reasons to want optionality multiply.

Where the squeeze lands

Production numbers add nuance: NiCE's telemetry shows Anthropic's Opus 4.7 takes 4+ seconds to first token and costs 18.5x more than Gemini 3 Flash on comparable real-time workloads — model selection is now a workload-specific decision, not a brand decision. And Atlassian is already weaponizing this, selling model-agnostic orchestration on data-sovereignty grounds: 'I don't want all my company's information going to Anthropic.' Every foundation lab moving up the stack into agent products will get squeezed by the SaaS incumbents it depends on for distribution.

When a rocket company buys a $60B coding startup while the leader's chatbot share drops below 50%, vendor concentration is a board risk, not an engineering decision.