40% fewer tokens, and a $6.7bn fab already sold out

Today's two digests cover the same week and point in opposite directions. One records three model releases whose headline claim is doing the same work with less. The other records a fab that sold out before volume production, a 300MW data centre cluster breaking ground, and $3.36 billion raised against compute that does not exist yet.

Less per task

Fireworks Research put out Ember-1 as a two-week research preview, trained to match Kimi K3's quality on about 40% fewer tokens, scoring 82.0 against K3-max's 80.9 on Terminal Bench 2.1 and 92.2 against 93.2 on SWE-Bench Verified, with two customers running it in production measuring about 35% fewer tokens per task.

Nvidia released Nemotron 3 Diarization, a 100 million parameter model with freely available weights that separates up to eight speakers in recorded or live audio and tops the VoiceArena diarization benchmark with a 14.72% error rate against 19.3% for the next best system.

Z.ai said its open-weight GLM-5.2 benchmarks between Anthropic's Opus 4.7 and 4.8 on the hardest agentic and long-horizon tasks.

Three different claims: fewer tokens for the same answer, a small model that wins a benchmark outright, open weights inside the frontier's range. Read together, they say the cost of a unit of useful model output fell.

More capacity

The capacity items from the same week read as though none of that happened.

Vanguard International Semiconductor inaugurated its $6.7 billion VSMC plant in Tampines on September 28 and is already evaluating a second Singapore fab: the first plant's roughly 44,000 wafers a month sold out ahead of volume production, and full utilisation moved forward to 2028 from 2029.

Hyundai Engineering won a 870 billion won contract to build the first 60MW of SGC Group's 300MW cluster in Gunsan, starting in October, with priority negotiation rights expected on the remaining 240MW.

China's Ministry of Industry and Information Technology asked Alibaba, ByteDance and other firms to report how many RTX Pro 5500 workstation GPUs they intend to buy. ByteDance alone is weighing roughly 1 million units against Nvidia's target of about 500,000 a quarter for the entire Chinese market: one buyer at twice the quarterly supply.

And **Nscale** raised $3.36 billion in convertible loan notes led by Third Point, a $2.36 billion tranche at closing plus a $1 billion Nvidia commitment due in mid-November, with the notes converting automatically on IPO.

The gap is the argument

Nscale is where the two columns meet. It signed a $1.4 billion GPU-backed term loan in February, raised $2 billion in March at a $14.6 billion valuation, and has about 200,000 Nvidia GB300 GPUs contracted to Microsoft. This week's convertible is money raised on the premise that demand for compute outruns every efficiency gain shipped against it.

Nothing in today's record settles whether that premise holds. What it does show is that the two sets of numbers are measured differently and nobody is reconciling them. The efficiency figures are per task: 40% fewer tokens, 100 million parameters, one model's weights in the open. The capacity figures are aggregate: wafers a month, megawatts, units a quarter, dollars raised. A model that needs 35% fewer tokens lowers no bill if the number of tasks rises faster, and the buyers signing 300MW contracts have read the same benchmarks as everyone else.

One item sits on the seam. Apple is evaluating a return to the commercial server market with a product aimed at AI inference, built on its own M-series chips and potentially using Nvidia's NVLink Fusion interconnect. Inference is the load that all of this week's efficiency work is meant to shrink. It is also the load Apple would be building new hardware to serve.