Welcome to AI Decoded, Fast Company’s weekly newsletter that breaks down the most important news in the world of AI. I’m Mark Sullivan, a senior writer at Fast Company, covering emerging tech, AI, and tech policy.
This week, I’m focusing on Elon Musk’s decision to lease the computing capacity at SpaceX’s Colossus 1 data center to Anthropic. I also look at what a new Atlantic exposé on David Sacks says about Silicon Valley’s alliance with Trump, and a benchmark that’s stumping top AI coding agents.
Sign up to receive this newsletter every week via email here. And if you have comments on this issue and/or ideas for future ones, drop me a line at sullivan@fastcompany.com, and follow me on X (formerly Twitter) @thesullivan.
Why Grok is selling compute to Anthropic
While everybody else in the AI space scrambles to lock down computing power, xAI’s Grok models are apparently being used so little relative to peers that the company can sell off the capacity of entire data centers, “colossal” ones at that.
Anthropic said Tuesday it had signed an agreement with SpaceX to use all the computing capacity in SpaceX’s Colossus 1 data center in Memphis. (SpaceX owns xAI.) The deal will give Anthropic access to more than 300 megawatts of computing capacity, or more than 220,000 NVIDIA GPUs. Anthropic says the additional capacity will be used to serve its Claude Pro ($20 per month) and Claude Max ($100 to $200 per month) subscribers.
SpaceX CEO Elon Musk says he gave his much-sought moral stamp of approval to Anthropic. “By way of background for those who care, I spent a lot of time last week with senior members of the Anthropic team to understand what they do to ensure Claude is good for humanity and was impressed,” Musk said in an X post. “Everyone I met was highly competent and cared a great deal about doing the right thing. No one set off my evil detector.”
Musk says xAI had already shifted its training workloads to Colossus 2, freeing up Colossus 1 for Anthropic’s use. Anthropic says it will use the facility primarily for inference, or the processing required to respond to user prompts in real time.
The partnership could eventually extend beyond Earth. Anthropic says it has also been discussing plans with Musk and SpaceX to develop multiple gigawatts of orbital AI compute capacity. Space-based AI data centers hold obvious appeal because the cost of cooling servers would essentially disappear. But major technical hurdles remain, especially around reliably transmitting massive amounts of data between orbiting infrastructure and Earth.
Musk’s willingness to arm Anthropic with vital computing power may also have something to do with his hatred of Anthropic rival OpenAI, and his dislike of OpenAI founder Sam Altman. Musk sued OpenAI, claiming the company’s leadership betrayed its original nonprofit mission to develop AGI for the benefit of humanity rather than for profit.
Trump’s bargain with Silicon Valley on AI may be weakening
The Atlantic’s George Packer, in a new article about former White House “crypto and AI czar” David Sacks, sheds more light on how and why Sacks and other Valley elites went full MAGA before the 2024 election. Now there are signs that the main thing Silicon Valley wanted in exchange for its support may be in jeopardy.
Silicon Valley’s preferred version of its MAGA conversion story is that influential VC Marc Andreessen met with representatives of the Biden administration and was told the administration intended to heavily regulate AI so that only a few big AI labs, and no startups, would be able to comply and survive. Andreessen said Biden wanted to “nationalize or destroy” Silicon Valley. He said Biden wanted to kill the entire cryptocurrency industry. He said he and his partner Ben Horowitz decided to support MAGA right after that meeting.
Biden officials dispute Andreessen’s account of what was said. But Andreessen’s version was enough to set a broader shift in motion among tech elites. Sacks held a fundraiser for Donald Trump in June 2024 in San Francisco’s wealthy Pacific Heights neighborhood. After talking with Trump at the event and on the All-In podcast, Sacks said: “All of his instincts are Let’s empower the private sector; let’s cut regulations; let’s make taxes reasonable; let’s get the smartest people in the country; let’s have peace deals; let’s have growth.”
What Sacks and others were really after was a promise of AI deregulation and more tax cuts. They got the tax cuts, and so far the Trump administration has worked hard to stifle government investigations or regulations targeting the tech industry. Some states have passed laws requiring government oversight, but the administration has been trying to preempt such laws or challenge them in court.
Packer suggests that Sacks, Andreessen, Horowitz, and other Valley elites may also share something in common with much of MAGA: They are white men witnessing a loss of status in society. “Andreessen was willing to pay high taxes and support liberal causes and candidates as long as he was regarded as a hero,” Packer writes.
But Silicon Valley’s fall from grace is not the fault of Democrats, Biden, or “wokism”; it’s the result of government and society slowly realizing that many Silicon Valley elites are not actually driven by idealistic notions of “making the world better.” Instead, they’ve repeatedly shown a willingness to unleash technologies they know may be harmful. The clearest example is Meta, which the government largely allowed to regulate itself while shielding it from many user lawsuits through Section 230, only to watch social media platforms contribute to disinformation, political polarization, and harms to children.
But nothing is permanent with Trump, as so many others have found out, and agreements that no longer provide immediate value can be quickly abandoned.
The White House announced this week that it’s considering a requirement that government officials “vet” new AI models before they can be released. Team Trump was apparently spooked by two things. An AI model from a company it recently declared a supply-chain risk, Anthropic, developed a model called Mythos that can identify software vulnerabilities at scale and devise ways to exploit them. Meanwhile, backlash against the tech industry’s massive data center buildout is becoming increasingly unpopular with parts of the MAGA base and could become a major GOP liability in the midterms.
Maybe tech elites and MAGA don’t mix quite as well as either side once thought.
Meet the new benchmark that’s soundly defeating coding agents
Perhaps the most consequential application of generative AI models so far has been software engineering, where agents generate code and increasingly make high-level architectural decisions. But how do we tell how good an AI software engineer really is? Until now, the industry has largely relied on benchmark tests such as SWE-Bench, which evaluate models on relatively well-defined tasks like fixing bugs or implementing a single feature. Now the developers behind SWE-Bench have introduced a much harder test called ProgramBench.
The benchmark is difficult because the AI agent has to reason strategically about the optimal architecture and programming language needed to reproduce the performance of each of the 200 test programs. Once an agent finishes building a codebase, the benchmark runs roughly 248,000 tests to measure how closely the recreated software matches the original behavior.
So far, all of the major models tested on ProgramBench, including Anthropic’s Claude Opus 4.7, Google’s Gemini 3 Pro, and OpenAI’s GPT-5.4, have scored big fat zeros. In other words, none were able to fully complete the test builds. Several models, however, were able to complete portions of them.
The results suggest that current AI coding tools still are not advanced enough to make the kinds of architectural and systems-level decisions human software engineers routinely make when turning an idea into working software. The findings may also indicate that AI agents still struggle to apply abstract principles learned during training to entirely novel problems.
More AI coverage from Fast Company:
- How a Texas vegan cheese-maker used Claude and Manus to fight back against a big shipping company
- AI power users are pulling away from everyone else, Microsoft says
- AI labels were supposed to help users spot fakes. Here’s why they’re failing
- OpenAI’s trillion-dollar AI bet is a study in ‘riskmaxxing’
Want exclusive reporting and trend analysis on technology, business innovation, future of work, and design? Sign up for Fast Company Premium.