Three things broke your AI strategy this week — name them out loud

Pick the week's most consequential signal and most people will say Musk v. Altman, which begins jury selection Monday with $100B in damages on the table and Nadella on the witness list. They'd be wrong. The trial is a binary event with a wide outcome distribution and a year of appeals built in. The things that already happened are the ones that change what you ship next quarter.

Three of them, stacked.

DeepSeek V4 ran natively on Huawei Ascend, and the export-control thesis is over

V4 Pro and V4 Flash shipped this week — 1.6T and 284B parameters respectively, MIT-licensed base weights, 1M token context, and inference running on Huawei's CANN stack with no NVIDIA in the loop. V4 Flash is priced at $0.14 per million input tokens. DeepSeek has publicly stated pricing falls further when Ascend 950 supernodes deploy in H2 2026.

Four of the five top open-weight models on current rankings are now Chinese: Kimi K2.6, V4 Pro, GLM-5.1, Qwen 3.6. All MIT-licensed. All with full technical reports. The open frontier is a Chinese-led market.

The catch is real and it matters: V4 Pro hallucinates 94% of the time on AA-Omniscience, V4 Flash 96%. It is simultaneously the #1 open agentic model on GDPval-AA and the worst factual model in its weight class. This is not a contradiction — it's a divergence. Capability and reliability are now separate axes, and your eval harness probably scores them as one.

The right read isn't "swap to V4 to save money." It's that the floor on inference cost just dropped through the basement on hardware the US can't sanction, and the model is excellent at executing plans built on fabricated facts. That's a routing problem and a verification problem, not a procurement problem.

Anthropic proved that tiered agents quietly extract value from weaker counterparties

Project Deal: 69 Anthropic employees, 186 real Slack negotiations over a week, ~$4,000 transacted. Some employees got Opus-backed agents, others Haiku. Opus sellers earned more. Opus buyers paid less. The Haiku side rated the deal's fairness identically to the Opus side. They had no idea they were losing.

Anthropic called it an "uncomfortable implication." That's generous. If your product tiers AI capability by pricing plan — and most do — and any agent-mediated workflow involves users transacting with each other, you have built systematic, undetectable wealth transfer between segments. Marketplaces. CRM negotiation features. Procurement automation. Job matching. The disadvantaged party can't churn over an asymmetry they can't perceive.

The sample is small and the stakes were play money. Treat it as directional, not statistical. The direction is alarming enough.

The move isn't to standardize everyone to your top tier. It's to identify which workflows are fairness-sensitive — the ones where agents on different tiers interact across the same transaction — and force model parity there, while differentiating on volume, latency, or non-adversarial features. If you can't standardize, instrument outcomes by tier and run the bias test yourself before someone else publishes it.

Microsoft is giving every Windows user a 35-day pause button, repeatable, with no cap

This is the operational story nobody outside security teams is tracking. The new feature lets users defer Windows updates 35 days at a time, indefinitely. Mean time-to-exploit for newly disclosed CVEs is now around 20 hours. The math is a 42x mismatch between user-controlled deferral and adversary weaponization.

Unless your MDM or GPO policy explicitly overrides this behavior, every endpoint user holds veto power over your patch SLA. Combine that with the serial-to-Ethernet converter RCE and auth-bypass disclosures hitting RTUs, PLCs, PoS systems, and bedside monitors this week — devices that often don't appear in standard asset inventories — and your patch surface fractured along two axes simultaneously.

Verify your group policy blocks user-initiated pause beyond your SLA. Test against current and upcoming Windows builds. Document it for your next audit. This is a one-sprint task and the cost of skipping it compounds the moment users discover the button exists.

What to actually do this week

The through-line across all three: the assumptions your stack was built on — that frontier models live behind a small set of US APIs, that capability tiers are an internal pricing decision, that patch deployment is something IT controls — are simultaneously false now. Not in eighteen months. This week.

One concrete operator move, ranked by leverage:

Run a fairness audit on every agent-mediated feature in your product. Pull last quarter's outcome data — win rates, prices paid, deals closed, matches accepted — and segment by user plan tier. If the delta is statistically significant and economically meaningful, you have a Project Deal problem in production. Standardize the model tier on those workflows before the next release.

If you have an hour after that, write the GPO that locks Windows update deferral to your SLA, and benchmark V4 Flash against your current routing on a representative slice of your workload — total cost per completed task, not per token, with hallucination rate measured on your domain.

The Musk trial will resolve when it resolves. The three things that already happened won't wait.