Microsoft Pulls Claude Code From Its Own Engineers After DeepSeek Locks In a Discount

01Trump labeled his own AI safety order an "innovation blocker" after the CEOs declined to attend the signing

The signing event was on the calendar. An executive order mandating baseline safety testing for frontier AI systems was drafted. Then top AI firm CEOs declined to attend, and the order Trump was prepared to sign became, in his own words, an "innovation blocker."

According to Ars Technica, Trump scrapped the signing this week. The president who had positioned himself as ready to impose federal guardrails on AI development reversed in days, blaming the same companies whose absence triggered the cancellation. No replacement framework was offered, no revised draft, no timeline. Federal AI pre-release testing rules, the kind that would have set a floor for what "tested before release" means, now have no path forward. The status quo holds: voluntary commitments, company self-reports, no mandatory disclosure of what evaluations were run before deployment.

The industry won. What it won is harder to define.

The Verge reported the same week that hackers have moved past prompt injection and are now exploiting the "personalities" chatbot makers ship as features. First-generation chatbot attacks, the column noted, were a "laughably simple" affair. Current attacks target the persona layer: the tuned character traits, refusal patterns, and conversational style companies use to differentiate their models. Attackers probe what tone breaks the guardrails, then write inputs that match.

In a TechCrunch column this week, the publication's blunt summary was that everyone is figuring out AI security on the fly, Google included. The company makes its own models, runs a cloud platform that hosts other companies' models, and publishes some of the most-cited AI safety research. If Google is improvising at this scale, the baseline the abandoned executive order would have established looks even more conspicuous in absence.

Neither side has a settled answer for the threat model the Verge described. Trump's order would not have mandated specific defenses against persona attacks; the CEOs who skipped the signing have not published one either. What the cancellation removes is the requirement to disclose what was tested at all.

Trump walked away without naming a date to revisit. The persona attacks documented this week were not waiting for federal policy.

02Two-thirds of an AI chip's cost is now memory. Samsung's memory engineers just turned that into $340,000 bonuses

A new analysis from Epoch AI puts memory at roughly two-thirds of the component cost of a modern AI accelerator. That share was much smaller when GPU compute first started scaling. The logic die, which gets most of the press attention, is no longer where most of the bill sits.

The same economics showed up in a labor negotiation this month. Samsung's memory chip employees had threatened an 18-day strike. They secured a tentative deal making some workers eligible for average annual bonuses of about $340,000, according to The Verge. The walkout had hinged on the company's bonus cap for semiconductor division staff, following what the report described as a substantial rise in the unit's earnings.

Read together, the two data points describe one mechanism. Customers buying H100-class and GB200-class accelerators are paying most of that bill to the memory suppliers stacked onto each package. That same revenue gave Samsung's semiconductor unit room to settle a strike threat with one of the largest bonus packages the industry has seen.

It also reframes the compute spending of model companies. A meaningful share of every multibillion-dollar GPU order now flows past the logic-die manufacturer and into HBM production lines that are already capacity-constrained. Whoever controls those lines can set price and absorb labor costs without giving up margin. The customers absorbing that pass-through are the same firms racing to lock in training capacity.

The bonus settlement is one consequence. A second is that a wage floor set at the largest memory supplier feeds into the cost stack of every accelerator its competitors ship. They recruit from the same engineering pool. AI buyers pay the difference.

Samsung's negotiation did not change the supply picture. It confirmed that the people working the bottleneck know exactly how scarce their output is.

03Microsoft revokes Claude Code from its own engineers as DeepSeek locks in its discount

Microsoft started canceling Claude Code licenses for its own employees this month, The Verge reported, citing internal communications. The cuts hit engineering teams across the company, removing access to Anthropic's command-line coding agent that some Microsoft developers had adopted on their own. Affected employees were told to switch to GitHub Copilot or Microsoft's internal coding tools.

The move came from OpenAI's largest backer. Microsoft has poured billions into OpenAI and ships Copilot products built on those models. Pulling a rival lab's tool off internal machines signals that Microsoft will no longer subsidize Anthropic's footprint inside its engineering org, even where individual developers prefer it.

DeepSeek spent the same week moving in the opposite direction. The Chinese lab launched its V4 Pro coding model with a promotional discount it had framed as a limited-time rate. A quiet update to DeepSeek's public API documentation made that discount the standard price. The countdown is gone.

For working developers, the two updates land in different parts of the market. Inside Microsoft-managed environments, the AI coding tool they can install just narrowed by one. Outside those walls, DeepSeek's V4 Pro discount stops being a deadline-bound promotion and becomes the baseline price. Indie developers and small teams paying per-token now have a permanent low-cost option to weigh against Claude or Copilot.

What changed is who decides which tools sit on the developer's machine. Microsoft is actively pruning AI coding tools its engineers can run, not only promoting Copilot. DeepSeek no longer needs a promotional window to keep developers on V4 Pro. The price that pulled them in becomes the price that keeps them.

The Verge did not say how many Microsoft developers lost access, or whether the cancellations cover all divisions or specific teams. DeepSeek did not announce the pricing change beyond the documentation update.