Monday's OpenAI trial is the loudest signal, not the biggest one

Jury selection in Musk v. Altman starts Monday, April 28. Musk is seeking $100B+ in damages, the removal of Altman and Brockman, reversal of OpenAI's for-profit conversion, and has named Microsoft as co-defendant. Nadella testifies. Altman testifies. Brockman testifies. A litigation observer called it "the Hindenburg landing on the Titanic."

That's the cover story, and it's a real one — a partial Musk win freezes OpenAI's IPO, creates governance uncertainty rippling through every Copilot dependency, and hands Anthropic an enterprise migration tailwind Google just spent $40B pre-positioning for. If your production stack goes through OpenAI or Azure OpenAI, you need a vendor abstraction layer live before opening arguments, not after the first bad testimony day.

Yes, but — the base rate for these outcomes says Altman wins or settles, and the market has been pricing OpenAI risk in the open for eighteen months. Fair. The point isn't that the trial breaks OpenAI; it's that the discovery process forces every AI-dependent org to answer a question they've been deferring: what happens to our roadmap if our primary model provider becomes uninvestable for six months? The trial is the deadline, not the disaster.

And it's not even the most important thing that happened this week.

The model layer just commoditized from below

DeepSeek V4 shipped running natively on Huawei Ascend chips. Not as a port, not as a benchmark stunt — as the deployment target, with a public roadmap tied to Ascend 950 supernodes in H2 2026. V4 Flash prices at $0.14 per million input tokens under MIT license. Chinese labs now hold four of the top five open-weight positions: Kimi K2.6, DeepSeek V4, GLM-5.1, Qwen 3.6.

The US export control thesis — restrict NVIDIA, keep the frontier American — is empirically dead. You don't have to like that conclusion to plan around it.

The catch is the one everyone should be citing and almost nobody is: V4 Pro scores #1 among open-weight models on the GDPval-AA agentic benchmark while hallucinating on 94% of factual queries in AA-Omniscience. Flash is worse at 96%. A model that can execute a ten-step plan flawlessly while confidently inventing the facts the plan was built on is a genuinely new failure mode. It's not the old "hallucination problem." It's capability and reliability decoupling on separate axes, and your single-score leaderboard is now actively lying to you.

This is where the operator move lives. Split your evaluation harness into two axes this sprint: agentic capability (can it do the thing) and factual reliability (does it know the thing). Measure them independently on your actual workloads. Route accordingly — cheap open-weight for execution and codegen, frontier API plus retrieval for anything where a wrong fact costs money.

And if you still price frontier API access as the default, pull the numbers. GPT-5.5 launched at 2x prior pricing. V4 Flash serves at roughly one-fifteenth of that. GPT-5.5 does emit 45–56% fewer output tokens per task, which claws some of the gap back on total cost per completed job — but only if you actually measure per-task cost instead of per-token price. Most teams don't. Most teams should, by end of next week.

Anthropic's Project Deal is the finding that should keep product leaders up

Sixty-nine Anthropic employees. 186 real transactions on Slack over a week. Some got Opus agents, some got Haiku. Opus sellers earned more. Opus buyers paid less. The losing side rated deal fairness identically to the winning side. They had no idea they'd been beaten.

Sample is small, stakes were play money, and the effect sizes weren't published — treat the magnitude as directional. The mechanism is not directional. Any product tiering AI capability by pricing plan across an adversarial workflow — marketplaces, procurement, matching, negotiation, recommendation — is building systematic economic asymmetry that the disadvantaged party structurally cannot detect and therefore cannot complain about. That's not a UX problem. That's a category of quiet product harm that regulators haven't named yet but will.

Audit every agent-mediated workflow this sprint. If two users on different plans interact through your AI, either standardize the model tier for that interaction or make the tier disclosed on both sides. There is no third option that survives a journalist finding out.

The infrastructure story nobody's pricing

Samsung's mobile chief warned of a first-ever net smartphone loss in 2026 — not from weak sales but because AI is eating global memory supply. One NVIDIA Vera server consumes RAM equivalent to 4,600 Galaxy S26 Ultras. Oracle's ~$300B data center push is straining Wall Street's ability to syndicate the debt. Twelve-plus US states are considering data center construction moratoriums.

Compute doesn't get cheaper from here for eighteen to thirty-six months. Any 2026–2027 plan that assumes falling inference costs on frontier APIs needs a second scenario where memory-driven hardware costs rise 20–30% and permitting friction slows capacity additions. Model both. Present both to whoever owns the budget.

What to do this week

One action per role, actually specific:

If you own infra: stand up a model-provider abstraction (LiteLLM, Portkey, or a thin gateway you own) in front of every production LLM call, and get baseline evals on Claude and one open-weight running against your top three workloads before Monday's opening arguments.

If you own product: pull the list of every AI-powered feature where two users interact through the model, and put model-tier parity on the design review checklist by Friday.

If you own security: verify your MDM policies block the Windows infinite update-pause behavior before it ships, and inventory serial-to-Ethernet converters across your OT footprint. Mean time-to-exploit is 20 hours. Your users should not hold a 35-day veto over that number.

If you own the P&L: rerun Q3 inference cost projections at GPT-5.5's new pricing, then rerun them again assuming a 60/40 split routed to V4 Flash or Kimi K2.6 for execution tasks. The delta is where your routing layer pays for itself.

The trial gets the headlines. The rest of the week is the actual work.