The trust question. The buyers stopped asking how good the model is. They started asking who can be trusted with the data.

Z.ai's GLM-5.2 — open-weight, downloadable, runnable on the customer's own machines — priced at roughly one-sixth of closed frontier models from Anthropic and OpenAI, according to Reuters. Snowflake's CEO and named venture investors praised the quality on the record — endorsement, not a production win.

Cheap tokens are not cheap tasks. Cursor's benchmark on real coding work puts Anthropic's Opus 4.7 at about $11 per completed task — and Cursor's own Composer 2.5 within two points at 55 cents. The buyer pays for finished work, not the per-token sticker.

Picture a developer routing a coding job through OpenRouter because the sticker is one-sixth of Anthropic or OpenAI. That is not the same as a regulated enterprise signing a production contract — and it is not the same as running the model inside your own walls. Snowflake has the compute, the security team, and the sandbox. A twenty-person startup does not. It rents through a router, and the Chinese origin risk stays live.

Alex Karp went on CNBC that week, stating enterprises are being harvested for their IP and data, not just overcharged. Played out in the media for a reason.

For most buyers the purchase still comes down to two numbers: cost and trust. A Chinese open-weight model clears the second only behind your own gates — an engineering problem for the few who can sandbox it, and perhaps a window for an organisation that can optimise for both.

Picture an enterprise buyer who has rebuilt their workflows around a single AI vendor. Signal 007 left the fork open: bet the company on one lab, or deploy in the pockets where the payoff is obvious and keep your options open. Microsoft — the king of diversification and capture — named the second path.

Frontier Company — an operating unit inside Microsoft, not a spin-out — with $2.5 billion and 6,000 engineers to help customers wire in AI tools, including open-source models, against their own data. Unilever, Novo Nordisk, and the London Stock Exchange Group were among the early names. The stated design principle: reduce lock-in.

Microsoft is OpenAI's largest backer. It does not lead the frontier on model capability. It does lead on decades inside enterprise systems. If the contest is whose model is smartest, Microsoft loses to the labs. If the contest is whom do you trust inside your systems, Microsoft has been building that answer for thirty years. The launch is a vote for which contest comes next.

Meta spent two years on an Ohio campus spanning roughly a thousand football fields, betting the model would justify the bill. In early July it began renting spare capacity and selling API access to its models. CoreWeave and Nebius, both AI compute suppliers, sold off on the news, as Meta is now the third lab — SpaceX and Anthropic before it — to come for their share.

You rent out spare racks when the return is in the electricity and the chips — not the model running on them. Meta made the infrastructure the product. The Signal 002 thread — spend to win AI becoming spend and make it pay — moved one notch deeper.

Trust requires stable access. For two and a half weeks, an enterprise team's best model had stopped answering. On 30 June the Commerce Department withdrew export controls on Anthropic's Fable and Mythos. Fable returned broadly; Mythos to trusted US organisations first. The reprieve came with safety commitments whose details remain undisclosed. OpenAI delayed GPT-5.6 at the government's request in the same window.

Washington's stated reason is national security. In the buyer's chair it is availability risk: a component that can vanish for two weeks. The same window has also carried speculation — not verified here — that vendors swap models when throttled, or route by task while the buyer still sees the model they selected. The invoice lands on the model you chose, not necessarily the model that did the work.

Alibaba banned staff use of Anthropic's Claude Code from 10 July, ordering employees onto Qoder. Security is Alibaba's stated reason — an allegation of hidden code tracking Chinese users, reported as an allegation, not a finding. Weeks after Washington gated China's access to America's best models, China's biggest cloud company gated an American lab's tools. Trust is the weapon on every side.

We do not know what changed in the Washington deal, or what Alibaba found. We note both.

While the stack fought over trust, PJM — the grid serving 67 million people across the eastern US — spent the holiday weekend in emergency mode. A heat dome pushed demand toward a record last seen in 2006; generator outages and overloaded lines kept the system on alert for a third straight day. Spot wholesale prices in northern Virginia — the world's largest data-centre cluster — surged past $2,000 per megawatt-hour against roughly $40 in normal conditions. Con Edison cut power to 17,000 New York customers as equipment failed under the load. That was a weather week.

The slower-moving number is year-to-date. Reuters put wholesale power costs in PJM territory up 68% in the first five months of 2026, to roughly $40 billion; of the year-on-year increase, data centres accounted for about $3.8 billion. Demand is outrunning new supply, and data-centre load is large enough to show up in that accounting. Gas plants are being financed to close the gap — C$4.6 billion in Alberta tied to a hyperscale build, $1.75 billion from National Grid for a Texas plant aimed at Microsoft — but those projects switch on in 2028 and 2030. The emergency was this week. The megawatts are next decade.