When Compute Became the Main Event

A record-breaking open model out of China, a very public scramble for chips, and even Google quietly hitting pause.

This week's AI story wasn't a single jaw-dropping demo. It was about plumbing — who has the compute, who can afford it, and who is willing to admit their next model isn't ready. The releases that mattered said as much about electricity and capital as they did about clever architecture. Here are the developments from the past week that actually move the picture.

China ships the largest open model yet

On July 16, Beijing-based Moonshot AI unveiled Kimi K3, a mixture-of-experts model with a reported 2.8 trillion parameters and a one-million-token context window. It is available now through Moonshot's API, with open weights promised by July 27 under a permissive license — which, if it holds, would make K3 the largest openly downloadable model released to date.

Early hands-on tests and Moonshot's own benchmarks put Kimi K3 at or near the top of public coding leaderboards, in some cases ahead of leading US systems. Take the leaderboard bragging with a grain of salt — these numbers move around and are partly self-reported. But the direction of travel is hard to miss: the biggest open model in the world now comes from China, its API pricing lands in the same range as mid-tier Western models, and DeepSeek's next release is expected within days. The open-weight frontier is no longer a Western story.

The scramble for compute goes public

The more revealing news was about what it costs to run models like these. Anthropic is reportedly in early, preliminary talks to lease AI computing power from Meta — a deal floated at around $10 billion, and a striking one given the two companies compete directly. Days earlier, Anthropic added Tom Blomfield, the Monzo co-founder and Y Combinator partner, to a team focused squarely on compute. Blomfield framed the move bluntly: as models start improving themselves, the availability of compute becomes one of the hardest problems to solve.

Follow the money and you see the same pressure. US venture firms deployed a record $412.7 billion in the first half of 2026, and roughly 86 cents of every dollar — about $356 billion — went to AI companies, according to PitchBook. Much of that is being converted, directly or indirectly, into chips and data-center capacity. Compute, not ideas, has become the binding constraint, and everyone is buying their way around it.

Google taps the brakes on Gemini 3.5 Pro

Not every headline was a launch. Google has reportedly delayed the wider release of Gemini 3.5 Pro by months after internal testing showed it falling short on coding and long-horizon reasoning; the company has not confirmed a date, a price, or a spec sheet. Reports describe a model that was effectively rebuilt rather than shipped.

It's a rare, public wobble from a lab that has otherwise shipped aggressively — and a healthy counterweight to the release-every-week mood. “Not ready” is a legitimate answer, even at Google's scale, and it beats pushing out something that stumbles in front of enterprise customers.

The labs start agreeing on the rules

Quieter but arguably more consequential: the CEOs of Google DeepMind, OpenAI, and Anthropic converged, in writing, on the idea that frontier models should face outside scrutiny before they reach the public — a real break from the industry's old self-reporting norm, as Axios reported. They still disagree on the how. Anthropic is backing state-level AI laws, while OpenAI is pushing for a single federal framework that would override them. For companies that long resisted binding oversight, publicly asking to be checked is a meaningful shift in posture — one worth noting as several of these labs head toward public markets.

---

The takeaway

The week's lesson wasn't any one model. It's that the frontier is increasingly gated by electricity, chips, and capital — and that even the leaders are bumping into the limits of what they can actually ship. The race is still on. It has just moved from the lab bench to the substation.