SoftBank's SB Energy delays a $50B IPO as investors balk
The first public price check on the AI data-center build-out came back too high for buyers, and a Russian lab put a from-scratch model on the open market the same day.
SoftBank's SB Energy has delayed the data-center IPO it had planned for this month after investors pushed back on a valuation above $50 billion, the New York Times reported Monday. The company filed its S-1 on September 1, targeting a Nasdaq listing and a raise of $5 billion to $7 billion, plus a Japanese retail tranche of up to $500 million, with pricing expected in the week of September 21. Its pitch rests on a backlog it values at roughly $439 billion, mostly data-center leases tied to OpenAI: 8.8 gigawatts contracted across Texas and Ohio, none of it operational yet, against $139 million of first-half revenue from its legacy renewables business and $3.21 billion of net losses over the same period. More than $170 billion of projected capital spending sits in front of all of it.
The mechanics of the delay are ordinary — when buyers won't clear at the price, banks cut the valuation or wait — but the timing is not. This was the reference listing for AI infrastructure: a SoftBank unit selling a stake priced on demand partly manufactured by its own parent, and it is also "substantially dependent" on a customer whose own listing has slipped. A lower valuation raises the cost of the equity that funds the build-out, which is why the pause lands as a market statement rather than a scheduling note. We flagged the dependency when the filing went in — SB Energy's IPO filing says it is 'substantially dependent' on OpenAI — and the public opposition that has been slowing projects is now showing up in the listings themselves, months after local opposition blocked $68B of US data centers in one quarter.
Yandex published an 80-billion-parameter model under Apache 2.0, and says it trained the whole thing from scratch. AliceAI-Foundation-80B-A3B-Base is a mixture-of-experts model that activates 3 billion parameters per token — the design that makes an 80B model cheap enough to serve — with a context window of 262,144 tokens and a training run Yandex puts at 18 trillion tokens. The weights went up on Hugging Face on September 18, with the documentation and announcement following Monday. It is a base model: no post-training, no alignment, not something to put in front of users, and Yandex says it will use it as the foundation for a reasoning system and the agent features in its Alice assistant.
The interesting part is the licensing, not the benchmark. Frontier labs build Russian-language capability they never release; Yandex has now handed out weights anyone can fine-tune and sell on, in a year when Moscow is writing a legal framework that separates "sovereign" models — domestic software, architecture and infrastructure — from merely national ones, with the formal status process not starting until March 2027. Experts quoted in Russian media are honest that this is a base model, smaller than the trillion-parameter frontier, so the sovereignty claim wins on paper before it wins on capability. Still, a from-scratch 80B release is an argument that the open-weight tier no longer depends on American or Chinese checkpoints as a starting point.
Linear rebuilt its CI pipeline because the AI coding agents it bought were making the queue the slowest part of shipping. The company's engineering blog details a year of work that started when the CTO filed an issue titled "CI costs are high": rewriting custom lint rules to use static analysis instead of type information, which dropped API lint time by 68% and full-repo lint by 55%, then attacking the setup cost every job pays — pulling the Postgres client into a CI base image, restricting package installs to the package under test, and booting containers from a schema snapshot instead of replaying migration history. Consolidating seven small checks into two jobs saved roughly 87,000 runner-minutes a month, about 11.8% of total CI usage.
The number worth stealing is the last one: with per-shard setup down to about 40 seconds from 110–140, Linear could afford to double its API test shards from four to eight, and the slowest shard fell from 5.25 minutes to 4.33. Agents now write the majority of the company's tests and it adds around 2,000 tests a week — the same curve every team that handed test-writing to a model is on. Nobody budgets for it, because a test generation is free and the runner minutes that validate it are not.
What to watch: whether SB Energy re-prices downward or waits for a better window — either one sets the benchmark other AI-infrastructure listings will be measured against.
If agent-written tests are the new CI bill, whose budget line should pay for them — the platform team or the team that generated them? Tell us in the comments.
