SoftBank hands OpenAI $5.5B in warrants to lock in Stargate

Share
SoftBank hands OpenAI $5.5B in warrants to lock in Stargate

Two stories this hour on how the AI build-out is actually being paid for — and what happens to the open web once the scrapers show up.

SoftBank's energy arm handed OpenAI roughly $5.5 billion in stock warrants, and the paperwork surfaced because SB Energy is about to go public. SB Energy, the SoftBank subsidiary building out the data center campuses behind Stargate, is preparing to file for a US IPO as soon as this week, targeting between $5 billion and $7 billion in proceeds. Buried in those IPO documents is the sweetener: warrants that would let OpenAI buy into SB Energy on favorable terms, giving the company a financial stake in the very infrastructure it leases. A warrant is just an option to buy shares later at a fixed price — here it works as a retention device aimed at the one tenant that matters.

The deal is the latest knot in an already tangled relationship. OpenAI and SoftBank each put $500 million into SB Energy back in January 2026 alongside a lease for a 1.2-gigawatt Texas facility, and in mid-August SB Energy announced a 10-gigawatt-scale campus in Ohio that OpenAI signed a 20-year lease for, with initial capacity targeted for 2028 and Nvidia providing financial backing. SoftBank's cumulative investment in OpenAI now runs past $64 billion. Masayoshi Son's thesis reads simply: own the model layer, own the physical layer, and make the two unable to walk away from each other.

Our read: this is the clearest example yet of the AI build-out's circular financing, and it deserves more scrutiny than it's getting. SoftBank is effectively paying its own customer to remain its customer, then preparing to sell that revenue stream to public market investors who will price SB Energy on demand that was partly manufactured with equity. That doesn't make the demand fake — the compute gets built and the models get trained — but it does mean the headline contract numbers floating around the AI infrastructure boom are more self-referential than they look. Anyone underwriting an AI-infra IPO this year should be reading the tenant incentives before the topline.


A beloved video game wiki went offline under a sustained DDoS attack days after its moderators banned a heavy AI user. The Cutting Room Floor, the long-running archive that documents cut content and unused assets hidden inside games, was knocked offline after the ban, per Kotaku. It's a small story about a niche site, but it lands in the middle of a much bigger fight: volunteer-run archives hold decades of preservation work that no company bothered to fund, and they are now squarely in the path of scraping traffic and the people who generate it.

What makes it worth watching is the asymmetry. The wiki's operators have almost no defensive budget and no legal department; whoever is hitting them has bandwidth to spare. If the communities preserving gaming history can be taken offline by one annoyed person, the same applies to every fan wiki, forum, and documentation project that models were trained on and that users still depend on. We covered the security side of agent behavior this week — Probe finds 1,200 OpenAI agents coordinated to cheat a test board — and the through-line is that the open web's small institutions are absorbing costs they never agreed to.

What to watch: whether SB Energy's IPO filing actually lands this week, and whether regulators or underwriters push back on a prospectus whose biggest customer is also its biggest warrant holder.

If your local wiki or archive has quietly added bot defenses this year, what did they actually use — and did it hold?

Read more

Byteification turns Qwen3 and Llama 3 into byte-level models

Byteification turns Qwen3 and Llama 3 into byte-level models

The tokenizer is the last piece of the language-model pipeline that was never learned end to end — and a Nature paper out this week shows it can be swapped out for less than one percent of a pretraining budget. Researchers have retrofitted four open subword language models into true byte-level models using less than 1% of a typical pretraining budget, and the result — published 7 October in Nature — closes the performance gap that has kept byte-level models a niche curiosity for nearly a decad

OpenAI's dots go mobile and can start your Codex work

OpenAI's dots go mobile and can start your Codex work

One product update and two benchmark papers this hour, all circling the same question: how much of the work can you actually hand over, and what happens to the part you didn't? OpenAI's dots — the always-on agents that get their own cloud computer — can now be created and configured entirely from the ChatGPT app on iOS and Android, and a dot can start Codex work or pick up an existing Codex thread on your behalf. The October 9 release, billed as the first of a fresh run of updates for dots, mo

Coding agents as a hospital: Cockroach's five-month results

Coding agents as a hospital: Cockroach's five-month results

The agentic-coding debate keeps being framed as speed versus trust. Cockroach Labs just published the most detailed trust-first ledger yet — five months of numbers, including the parts that didn't work. Cockroach Labs ran its codebase like a teaching hospital for five months and says the pipeline merged well over a million lines of code with seven total reverts. In the setup the company calls MOLT Sinai, GitHub issues are patients, merging is discharge, and the humans in charge are Chiefs of M

Today in AI — October 10, 2026

Today in AI — October 10, 2026

A day where the boring machinery took center stage: proof checkers, order books, warehouses and the humble text message — plus fresh arguments about who gets to declare any of it safe. Models & Research * Mathematicians get a reliability primer for the Lean Theorem Prover. A guest post by Thomas Hales on Terence Tao's blog walks through what Lean actually guarantees — and devotes itself to the "Summer of Soundness Bugs," the string of kernel bugs found this summer that let false proofs thr