Lab notes.

Microsoft cut 4,800 roles and went out of its way to say AI didn't do it. The gap between what companies claim and what economists see is now the actual story, and this week produced unusually clean evidence for both sides.

A former Bitcoin miner just signed a $19B, 20-year lease with Anthropic, and Brookfield 5x'd its fuel-cell financing framework to $25B. The pacing item for the AI buildout isn't silicon. It's power, and the capital to secure it.

vLLM shipped 0.24.0, SGLang landed agent-written kernels, and NVIDIA claims DeepSeek V4 got 5x cheaper to serve on Blackwell in a month. A look at where the marginal cost-per-token wins are actually coming from.

In a single week, NVIDIA started financing the buyers of its own chips, Meta hinted it might sell compute instead of buying it, and Together AI raised $800M. The FLOPS aren't the story anymore. The balance sheet is.

In an investor Q&A, Sony's gaming chief said the next PlayStation aims for experiences 'beyond the living room.' It is not a product, a spec, or a date. It is a sentence, and the most honest thing to do is treat it as one.

Anthropic shipped its first public Mythos-class model on June 9. Three days later the US government ordered it switched off worldwide. Nineteen days after that, it came back. The whole episode is a landmark in how frontier AI meets the state, and it deserves a careful, evenhanded account.

No fireworks, and that was the point. Microsoft spent two days giving AI agents the hardware, context, guardrails, and rails to run in production, and shipped seven of its own models trained without OpenAI. The strategy matters more than any single box.

Microsoft's most powerful Surface ever runs on NVIDIA's RTX Spark, packs 128GB of unified memory and full CUDA, and pointedly drops the Copilot+ branding. The chip choice, and what it signals, is the whole story.

Forty-one days after Opus 4.7, Anthropic shipped a modest but real upgrade: parallel subagent workflows, user-facing effort control, cheaper fast mode, and a sharpened emphasis on honesty. The most interesting line in the release is one flagging a concern about the model itself.

New Surface Pro and Laptop for Business on Intel's Panther Lake, with a hardware privacy screen and a memory-crisis price bump. The quiet story is the deliberate split: Intel for the enterprise, Snapdragon for consumers.

One hundred announcements, a 24/7 personal agent, smart glasses, and Search's biggest change in decades. Google's most aggressive AI-everywhere push yet is genuinely impressive, and a fair look also notices how much of it is last year's promise with a heavier price tag.

The first genuinely new base model since GPT-5.0, tuned hard for agentic coding and computer use. The benchmarks are real but partial, the pricing debate misses the point, and the strategy underneath is a shift from selling capability to selling cost-per-task.