An open-weight model now sits third on the most-watched intelligence index, the layer that serves models just drew a $17.5 billion valuation, and Brussels is putting a draft legal definition under the word “sovereign”. The common thread: where a model runs is becoming as consequential — and as regulated — as which model you pick.
An open-weight model reaches third on the frontier composite
Artificial Analysis refreshed its Intelligence Index on 18 July, and across 165 tracked models the top of the board reads: Claude Fable 5 at 59.9%, GPT-5.6 Sol at 58.9%, and — the part worth pausing on — Moonshot AI’s Kimi K3 at 57.1%, third overall. The two models ahead of it are proprietary, metered, and served on US infrastructure. The one in third place is a 2.8-trillion-parameter open-weight release whose full checkpoints Moonshot has promised for 27 July. For the first time, the third-best general-intelligence score on the most-watched composite belongs to something an enterprise can download and run inside its own perimeter.
That reframes the choice more than it reorders the ranking. A meaningful gap between the best proprietary model and the best open one used to be the argument for renting; at 57 versus 60 it becomes an argument about control rather than capability. Kimi K3 gets there on a hybrid linear-attention design that activates 16 of 896 experts per token and claims roughly 2.5x the scaling efficiency of its predecessor — the sort of engineering that makes a three-trillion-parameter model tractable to serve at all. The caveat is the one we flagged on 17 July: the weights are still an announcement, not an artifact, until the 27 July drop actually lands. Treat the benchmark as real and the availability as pending.
The serving layer is now worth $17.5 billion
Fireworks AI raised a $1.505 billion Series D at a $17.5 billion valuation on 16 July, led by Atreides Management, Index Ventures and TCV with NVIDIA and Lightspeed among the participants. The figures that explain the round: annualized revenue past $1 billion, up fivefold year-on-year, on more than 40 trillion tokens served per day — nearly triple its prior run rate. More than 95% of those tokens come from models specialized on customers’ own data, not general-purpose frontier calls.
That mix is the story. The venture market has put a top-tier valuation not on a frontier lab but on the layer that serves other people’s models — increasingly open-weight ones — tuned on proprietary data. It is the same thesis that makes a sovereign inference platform viable: the durable value sits in running the right model well on your own terms, not in owning the single largest model. Once the serving layer is worth $17.5 billion, “which model” stops being the whole question and “where and by whom it runs” becomes the rest of it.
Europe moves to give “sovereign” a legal definition
The EU’s Cloud and AI Development Act is a Commission proposal adopted on 3 June — not yet law; it is now in the ordinary legislative procedure before Parliament and Council. What makes it worth watching anyway is the draft’s core move: turning digital sovereignty from a slogan into a procurement rulebook. It defines four assurance levels for cloud and AI services, from Level 1 (data processed and stored on EU-located infrastructure) through Level 3 (EU ownership and control, with personnel criteria) up to Level 4 — full supply-chain transparency and no third-country interference. Public-sector bodies would pick a level based on their own risk assessment, and providers would be recognised by Member States after an audit.
The Commission is explicit about why the framework exists: over-reliance on non-EU cloud providers is named as a risk to Europe’s digital autonomy — the same exposure the US CLOUD Act creates for regulated data. The tiers only bind once the regulation is adopted and its transition periods run, so this is a starting gun rather than a finish line. But the direction is set: if CADA passes in anything like its current form, “sovereign” gets a legal definition in Europe, and it is one that maps cleanly onto where, and by whom, infrastructure is owned.
Quick Hits
- Kimi K3’s weights are the 27 July date to watch — the 2.8T open-weight model is public in preview, but the downloadable MXFP4 checkpoints are what turn a benchmark into something deployable.
- Benchmarks are moving toward real work — Artificial Analysis now runs a GDPval-AA leaderboard scoring models on tasks drawn from 44 knowledge-work occupations, not just exam-style prompts.
- Brussels pairs the proposal with capacity — alongside the CADA draft, the EU AI Office’s 15 July frontier-AI findings and the Cybersecurity and AI Action Plan fund EU capability to evaluate frontier models before they reach the market, operational by 2027.
- Capital keeps chasing cheaper inference — Fireworks’ round is the largest of a wider week of enterprise-AI infrastructure financing aimed at serving specialized, open-weight models more cheaply than frontier APIs, per CNBC.