Reflection AI signed $7.3 billion of GPU rent; first model Beam used under 3% of it, weights ship free
Beam cost, at listed cloud rates, under 3% of the $7.3 billion of GPU rent Reflection has committed, trails the Chinese leaders on its own benchmark table, and ships free this month against a $25 billion valuation.
Vincent Jiang · 3 min read
A $25 billion lab ships its first model, free
Reflection AI finally has a model. Beam, unveiled on October 5, carries 501 billion total parameters and activates 23 billion per token, built for coding and agents 12. It runs on a waitlist today; later this month the weights go out free under an Apache 2.0 license 12.
$7.3 billion of GPU rent sits behind it
Reflection signed a $6.3 billion compute deal with SpaceX in June for Nvidia GB300 capacity at Colossus 2, then a $1 billion deal through cloud provider Nebius in July 3. Behind the launch sits a $25 billion pre-money valuation from June's round, with Nvidia, Sequoia and Citigroup on the cap table 34. If the flagship is free and cheap to serve, the rent risk runs to the counterparties on those deals, SpaceX and Nebius, and to holders of Reflection paper marked at that $25 billion.
The build is a rounding error on the bill
Beam's whole training run, on Reflection's own numbers, is small change against that commitment. Pretraining used 6,144 GB300s for under four weeks; reinforcement learning used 10,500 for four weeks 1. That is at most about 11.2 million GPU-hours. Clouds listed GB300 capacity at $3.02 to $18 per GPU-hour in late June 5, which prices the entire run between roughly $34 million and $201 million: under three cents of every committed dollar at the dearest listed rate. The $7.3 billion does not buy Beam. It buys models that do not exist yet, and Reflection says a larger successor is already training 13.
Built to bill less, trailing on its own scoreboard
Beam activates 23 billion parameters a token against 40 for GLM-5.2, 49 for DeepSeek V4 Pro and 104 for Kimi K3 26. Reflection claims three to four times less inference compute than GLM-5.2, an estimate it calls approximate, not measured 12.
Beam activates 23 billion parameters a token, under a quarter of Kimi K3's 104
Data
| Value | |
|---|---|
| Kimi K3 | 104B |
| DeepSeek V4 Pro | 49B |
| GLM-5.2 | 40B |
| Beam | 23B |
Its own table puts Beam sixth of eight on Terminal Bench v2.1 at 80.1, behind DeepSeek V4.1 Flash, Kimi K3 and GLM 5.3, and 44.4 against Flash's 74.2 on DeepSWE 1.
Beam is the best Western open model in its own table and still sixth of eight
Data
| Value | |
|---|---|
| DeepSeek V4.1 Flash | 90.6 |
| Kimi K3 | 88.3 |
| GLM 5.3 | 88.2 |
| Qwen 3.8 Max | 86.6 |
| GLM 5.2 | 81 |
| Beam | 80.1 |
| Inkling | 63.8 |
| Nemotron 3 Ultra | 56.4 |
No independent evaluation of Beam has been published 2.
The case for the bill
Sequoia partner Stephanie Zhan, who invested in Reflection, argues that what decides agents is compute per task, not compute per model 7. Artificial Analysis, testing on vendor access, calls Beam one of the most token-efficient open models it has seen 2. If capability per unit of compute is the product, the rent is the strategy, not the contradiction.
Weights this month, an IPO next month
Nvidia collects on every path: it sells the GB300s, holds a stake in the tenant, and convenes the Nemotron Coalition promoting open weights 36. Anthropic, expected to go public next month in perhaps the biggest IPO in history 7, will ask investors to price a closed franchise whose newest challenger works for nothing. Watch the independent benchmarks once the weights land 2.
Deepdive
AI-generated from this story and its cited sources. Not investment advice.
