Unravelling the Stargate spin
Transformer Weekly: DSIT's indecision, ever-shortening timelines, and DeepSeek's new model
Welcome to Transformer, your weekly briefing of what matters in AI. If you’ve been forwarded this email, click here to subscribe and receive future editions.
Housekeeping note: Transformer is taking next week off. We’ll be back on Feb 7 with a special edition from Paris, ahead of the AI Summit. (And if you’ll be in Paris and want to chat, drop me a line! shakeel@transformernews.ai).
Top stories
OpenAI, SoftBank and Oracle announced Stargate, which will spend $100b to build AI data centers this year, with an additional $400b over the next four years.
The project, announced at the White House on Tuesday, mostly builds on existing, widely-reported plans — and does not involve any US government spending (though President Trump said he’ll help the project build the energy capacity it’ll need).
At the announcement, Altman said “we couldn’t do this without you, Mr President”. But as the Washington Post reported, the project “was in the works for months before President Trump’s election victory — undermining … declarations that the venture was only possible because of Trump’s inauguration”.
OpenAI spokesperson Liz Bourgeois said: “This deal at this magnitude only came together following the election as a result of heightened investor enthusiasm in anticipation of a Trump administration.”
The FT has the neatest summary of the project I’ve seen:
“Despite the flashy announcement, Stargate has not yet secured the funding it requires, will receive no government financing and will only serve OpenAI once completed, the people familiar with the initiative have said.”
“[A] person close to the project said it was far from a fully developed plan: ‘They haven’t figured out the structure, they haven’t figured out the financing, they don’t have the money committed.’
SemiAnalysis (SA) has a superb, more detailed, breakdown of what’s really going on here. Some highlights:
“Only the first $100B has datacenter sites and power ready to go”.
“We believe this project is being measured on the Total Cost of Ownership, not Capital Expenditures.”
SA thinks all of the initial $100b is going to the Oracle-Crusoe-Lancium facility in Abilene, which broke ground last June.
The initial plan here was a 1GW facility; SA now thinks it’s going to have a total of 2.2GW of power supplied and a critical IT capacity of 1.8GW.
SA estimates the campus will house a previously-announced 100k GB200 cluster, as well as a 200k GB300 cluster and a 400k VR200 cluster.
As for financials: The Information reports that OpenAI is committing $19b for a 40% stake in the project, with SoftBank committing another $19b.
Oracle and MGX will reportedly provide another $7b, while “other investors and debt” will account for the rest.
Neither OpenAI nor SoftBank actually have that much in cash right now — so they’re going to have to raise it (in SoftBank’s case, likely by selling down some of its stake in Arm).
As for the other $400b … this all seems like more of a hope than a reality right now (though if AI progress goes as planned, I suspect no one will have trouble raising the funds).
The Information also has a great piece on how the deal came together.
In short, OpenAI was increasingly frustrated by Microsoft’s slow progress in building compute capacity (itself seemingly motivated by scepticism at Microsoft about AI progress).
Microsoft is now no longer OpenAI’s exclusive cloud partner, but it does still have “right of first refusal” on building capacity for OpenAI.
When Musk decided to ditch Oracle and Crusoe’s Abilene site, “Altman pounced” to secure it for himself.
Also notable is the much-discussed spat between Altman and Elon Musk, who immediately undermined the White House announcement by saying (seemingly correctly) that “they don’t actually have the money” (and suggesting that OpenAI and SoftBank were smoking crack).
Musk’s comments reportedly infuriated White House officials. Altman leveraged that tension by doubling down on sucking up to Trump, recasting his prior condemnations of the president by saying he “fell in the npc trap”.
But perhaps the most interesting response to Musk came from Satya Nadella: after Musk said that Nadella, unlike Altman, “definitely does have the money” for Microsoft’s data center projects, Nadella said “and all this money is not about hyping AI, but is about building useful things for the real world!”
Meanwhile, the UK government appears to be bottling it on AI regulation.
The AI Bill consultation is reportedly delayed, according to Politico, and officials are now scared to regulate in case doing so upsets Trump.
Apparently, UK officials are incapable of thinking for themselves, and are instead “[waiting] to see where the safety debate goes this year” before deciding what to do.
One notable nugget: while Peter Kyle previously said that the bill would include “a compulsion for pre-release of the models for testing by the relevant safety institute”, he’s recently stopped saying that.
I asked DSIT for comment on this; they gave the following non-statement:
“This government remains fully committed to bringing forward legislation which allows us to safely realise the enormous benefits of AI for years to come, and we will do so as Parliamentary timetables allow.
As you would expect, we are continuing to engage extensively to refine our proposals and will launch a public consultation in due course to ensure our approach is fit for purpose and can keep pace with the technology’s rapid development.”(I would suggest that one way to “keep pace with the technology’s rapid development” is to stop sitting around and actually do something — at the current pace, we’ll have AGI before the government’s bothered to do anything at all.)
Meanwhile, a new poll from the Centre for Long-Term Resilience and Public First found that 78% of the British public believe the government should regulate AI technologies, with 64% saying addressing extreme AI risks should be a top priority.
The discourse
Dario Amodei is increasingly bullish on short timelines:
“I am more confident than I have ever been at any previous time that we are very close to powerful capabilities.”
“I don't think it will be a whole bunch longer than [2027] when AI systems are better than humans at almost everything, and then eventually better than all humans at everything.”
“I think until about 3 to 6 months ago I had substantial uncertainty about it. I still do now, but that uncertainty is greatly reduced. I think that over the next 2 or 3 years I am relatively confident that we are indeed going to see models that … gradually get better than us at almost everything.”
Demis Hassabis thinks we’re very close to AGI, too:
“I think we’re closer and closer … I think we’re probably a handful of years away … “I would say probably like three to five years away.”
Brad Lightcap emphasised the pace of progress as well:
“We have gone from o1 to o3 in about three months. The iteration cycle is compressing. We’re already training the model that comes after o3: it looks like we’re going to see another big jump in capabilities.”
And Jake Sullivan thinks tackling AI risks must be the key priority for the next few years:
“I am a person who believes that we can seize the opportunities of AI. But to do so, we've got to manage the downside risks, and we have to be clear-eyed and real about those risks.”
On Transformer: I asked whether Elon Musk still cares about AI safety:
“If Musk genuinely believes there is a 10-20% chance of human extinction from AI, his failure to focus on reducing that risk represents a stunning abdication of responsibility. Political power is fleeting, and Musk has a once-in-a-lifetime opportunity to steer policy in a better direction. By his own lights, everything else Musk cares about — as important as it may be — pales in comparison to the importance of getting this right.”
Policy
As expected, Trump revoked Biden’s AI executive order. On Thursday, he implemented a new placeholder EO which:
Tells David Sacks, Michael Kratsios and Michael Waltz to develop an “AI Action Plan” within six months;
Orders a review of “all policies, directives, regulations, orders, and other actions taken pursuant” to the Biden EO.
Says the US “must develop AI systems that are free from ideological bias or engineered social agendas”;
And says “It is the policy of the United States to sustain and enhance America’s global AI dominance in order to promote human flourishing, economic competitiveness, and national security”.
Trump also said he’d use his energy emergency declaration to approve power station buildouts for AI data centers.
Rest of World has a great piece on how countries outside the West, including India, are very reluctant to regulate AI companies.
OpenAI told an Indian court that removing an Indian news agency’s content from ChatGPT's training data would violate its US legal obligations.
Influence
Monday’s inauguration was full of AI luminaries: Elon Musk, Mark Zuckerberg, Jeff Bezos and Sundar Pichai were all in the rotunda, while Sam Altman and Alexandr Wang were in the overflow room.
Wang took out a full-page WaPo ad saying that America must “win the AI war”, and do so by, among other things, “testing frontier AI models and systems to address unforeseen risks”.
Altman and Trump reportedly had a “lengthy” phone call last Friday, in which Trump was “particularly animated about building AI in the US instead of having it developed in China”.
OpenAI increased its lobbying spending to $1.76m in 2024.
Google’s Kent Walker wrote an op-ed in Fox News, calling on Trump to “make AI stand for American Innovation”.
CNAS put together a bunch of recommendations on how Trump could “promote and protect America’s AI advantage”.
AE Studio launched a campaign encouraging Trump to pursue an “AI Alignment Manhattan Project”.
Industry
DeepSeek released r1, an open-source reasoning model that performs very well at very low costs.
Nathan Lambert has an excellent and detailed explanation of how it works.
Zvi Mowshowitz has a very detailed analysis, too, including the implications for AI risk and policy.
It is, of course, heavily censored. And DeepSeek founder Liang Wenfeng met Chinese Premier Li Qiang this week.
The NYT profiled the company too.
Google reportedly invested an additional $1b in Anthropic, which is seemingly on track to raise a further $2b from other investors at a valuation of close to $60b.
Mark Zuckerberg said Meta will own 1.3m GPUs by the end of 2025, and that the company is raising 2025 capex spending to $60-65b.
ByteDance reportedly plans to spend over $12b on AI infrastructure in 2025, including $5.5b on Chinese chips and $6.8b on infrastructure outside of China.
Mukesh Ambani's Reliance Group announced plans to build a 3GW AI data center in India.
Huawei is reportedly making a push to displace Nvidia in the Chinese market for AI inference chips.
OpenAI is reportedly planning to value its non-profit’s stake in the company at around $30b. As Transformer has previously reported, this is arguably much less than it’s worth.
OpenAI launched Operator, its version of an AI agent that can use a browser to take actions.
It’s powered by a new model called “Computer-Using-Agent” which OpenAI says combines GPT-4o with “advanced reasoning through reinforcement learning”.
It scores 38.1% on the OSWorld agent evaluation — much better than the 14.9% Anthropic’s computer use model scored last year.
It’s also reportedly developing an advanced AI coding assistant aimed at replicating senior software engineers' skills.
Gemini on Android can now perform tasks across multiple apps in one prompt.
OpenAI announced a biology-focused model, in collaboration with the Altman-funded longevity startup Retro Biosciences.
Hugging Face released two very small multimodal AI models.
SK Hynix reported record profits, driven by AI demand, but investors seem to be spooked by its spending plans.
Scale AI competitor Invisible Technologies reportedly doubled its revenue to $134m last year.
Moves
Jeffrey Kessler is reportedly a favorite to lead the Bureau of Industry and Security.
Anduril promoted Christian Brose to president and chief strategy officer. Matthew Steckman is now president and chief business officer.
Abigail Wilson is now state policy manager at the Software and Information Industry Association.
Matt Pearl is CSIS’s new Strategic Technologies Program director.
Martín Soto is joining UK AISI as a technical advisor.
Best of the rest
Scale AI and the Center for AI Safety released “Humanity’s Last Exam”, an unfortunately-named AI benchmark intended to be much harder than anything else.
r1 currently has the best score of any tested AI model, with 9.4%. Dan Hendrycks said he thinks we might see scores of over 50% by the end of the year, though.
Relatedly: on Transformer, Lynette Bye wrote about how researchers think we might soon need to reevaluate how we safety-test AI systems, moving from “incompetence” evals to “countermeasure” evals. Read the full piece here.
The WSJ profiled Anthropic’s Frontier Red Team, which is recruiting.
OpenAI researchers found that giving reasoning models like o1 more inference-time compute can improve robustness to adversarial attacks.
New reporting revealed that Microsoft and Google both provided AI tools to the Israeli military in its war with Gaza.
BlueDot is launching a course for economists on the economics of transformative AI.
A report from Oxford Martin School recommended topics for US-China dialogues on AI safety and governance.
The AI Summer podcast continued its excellent guest lineup, with new episodes with Sam Hammond and Lennart Heim.
A new paper found that LLMs “have a form of intuitive self-awareness”.
LinkedIn was sued for allegedly sharing users' private messages without permission to train AI models.
People realised that OpenAI funded development of the FrontierMath benchmark, which Epoch AI didn’t initially disclose. OpenAI insists it didn’t train on the benchmark.
30% of game developers believe generative AI is having a negative impact on their industry.
The CIA has developed an AI chatbot to simulate world leaders.
Thanks for reading: have a great weekend, and see you in a couple weeks.