The Perpetual Promise Machine

The Colossus Lease

Start with the headline: Anthropic will rent the entire Colossus 1 data center from SpaceX. That is 220,000+ NVIDIA GPUs (a mix of H100, H200, and GB200 accelerators) and more than 300 megawatts of compute capacity. Anthropic says the capacity will improve service for its paid Claude Pro and Claude Max subscribers. SpaceX says it has "already moved training to Colossus 2."

That framing deserves scrutiny. When a company builds what it calls the world's largest supercomputer, brands it as the engine of artificial general intelligence, and then leases the entire facility to a direct competitor, the question is not "why is the competitor paying for it?" The question is: "Why isn't the builder using it?"

Musk's explanation on X was characteristically dramatic. He said he spent a week with Anthropic's senior team, that "no one set off my evil detector," and that SpaceX "reserves the right to reclaim the compute if their AI engages in actions that harm humanity." It is a useful narrative. It reframes what looks like a financial necessity as a philosophical concession. But the numbers underneath the narrative tell a very different story.

The 11% Supercomputer

In early May 2026, The Information reported that xAI's model flops utilization (MFU) across its GPU fleet was approximately 11%. To put that number in context: Meta achieves 43%. Google achieves 46%. Even OpenAI's early GPT-3 training runs, which were considered inefficient by today's standards, achieved 21% to 26%.

xAI operates roughly 550,000 GPUs across its Memphis and Colossus clusters, primarily H100 and H200 accelerators. At 11% utilization, that is the functional equivalent of 60,000 GPUs doing useful work. The other 490,000 are, in practical terms, generating heat.

xAI's own president, Michael Nicolls, described the utilization rate as "embarrassingly low" in an internal memo. He set a target of 50%, with no public timeframe for achieving it. The problem is not the hardware. It is the software stack: the distributed training architecture, the parallelization techniques, the data pipelines, and the GPU-to-GPU communication overhead. At scales of hundreds of thousands of GPUs, coordination problems compound. Idle time accumulates. Bottlenecks multiply.

This matters because the narrative around Colossus has always been about speed of construction, not quality of operation. Musk has repeatedly highlighted that the facility was built in 122 days and then doubled in 92 days. Those are impressive construction timelines. But building a supercomputer fast and running a supercomputer efficiently are two entirely different engineering challenges. The first is a logistics problem. The second is a software problem. And the software problem is where xAI's leadership has now departed.

The Anthropic lease starts to look less like generosity and more like the rational economic behavior of a company sitting on underutilized hardware and widening losses. If your fleet is running at 11% and your competitor will pay market rate to use it, leasing is not magnanimity. It is accounting.

The Talent Tells the Story

When Musk founded xAI in March 2023, he recruited 11 cofounders from the most accomplished AI research teams in the world: DeepMind, Google, OpenAI, Microsoft Research, and the University of Toronto. The founding team included Geoffrey Hinton's student Jimmy Ba, Google veteran Christian Szegedy, and chief engineer Igor Babuschkin from DeepMind and OpenAI.

By March 28, 2026, every single one of them had left.

The departures did not happen overnight. Szegedy left in early 2025. Kyle Kosic jumped to OpenAI in mid-2024. Babuschkin departed in August 2025 to launch an AI safety venture capital firm. But the cascade accelerated dramatically in February 2026, immediately after SpaceX's acquisition of xAI. Tony Wu, who led the reasoning team, announced his departure on February 10. Jimmy Ba, who led research and safety efforts, resigned within 24 hours. By mid-March, only two cofounders remained. Their departures on March 27 completed the sweep.

It was not just cofounders. A Fast Company review identified more than 80 people, including engineers and technical staff, who departed within the past year. The C-suite also hollowed out: xAI lost its general counsel, its chief financial officer (who departed after just three months), and its head of product engineering. Linda Yaccarino, who served as X's CEO, left in July 2025 and was never replaced.

In AI, talent is the product. Models do not improve because you buy more GPUs. They improve because researchers figure out how to use those GPUs more effectively. That is exactly the 11% utilization problem. And the people who were hired to solve it are the ones who left. The AI talent market in 2026 is the most competitive it has ever been. Meta has reportedly offered packages worth up to $300 million over four years to retain top researchers. Anthropic, OpenAI, and Google DeepMind are all expanding aggressively. The 11 researchers who walked away from xAI represent a concentration of expertise that any of those organizations would pay handsomely to acquire.

The Financial Reality

Internal financial documents reviewed by Bloomberg show that xAI recorded a net loss of $1.46 billion in the quarter ended September 2025, up from roughly $1 billion in Q1 2025. The operating loss for Q3 was $922 million. Over the first nine months of 2025, xAI spent approximately $7.8 billion in cash. That works out to roughly $28 million per day.

Against that burn rate, revenue was $107 million for the September quarter, nearly doubling from the prior quarter but still a rounding error against the losses. Total revenue for the first nine months was $208 million, with full-year 2025 projected at approximately $500 million. Bloomberg separately reported that xAI expected to lose $13 billion in 2025 on that $500 million in revenue. That is a 26-to-1 loss-to-revenue ratio.

For context, OpenAI was projecting roughly $12.7 billion in revenue for 2025. Both companies are losing money. But the scale of the gap between revenue and spending at xAI is in a category of its own.

Then came the merger. In February 2026, SpaceX acquired xAI in an all-stock deal that valued SpaceX at $1 trillion and xAI at $250 billion, creating a combined entity worth $1.25 trillion. That $250 billion valuation applies to a company doing approximately $500 million in annual revenue with accelerating losses and 11% GPU utilization. SpaceX itself reported revenue of $16 billion in 2025. Even adding xAI's contribution, the combined entity did not meaningfully exceed $20 billion in revenue. The IPO is targeting a valuation of $1.75 trillion.

The cross-entity capital flows add another layer. Tesla (TSLA) invested $2 billion in xAI and sold $430 million of Megapack battery storage to xAI in 2025 (costing Tesla $285 million to produce). xAI previously acquired X (formerly Twitter) in a $33 billion all-stock deal. Now SpaceX has acquired xAI. The money circulates between Musk entities with each transaction creating a new valuation that serves the next transaction. We covered this financial architecture in detail in The Musk Shell Game.

The FSD Playbook: Ten Years of "Next Year"

The xAI story becomes more revealing when placed alongside another Musk timeline: Tesla's (TSLA) Full Self-Driving (FSD) promises. The pattern is identical. The only thing that changes is the year.

December 2015: "We're going to end up with complete autonomy, and I think we will have complete autonomy in approximately two years."

October 2016: Tesla would demonstrate a fully autonomous drive from Los Angeles to Times Square "by the end of next year."

January 2017: When asked about FSD timeline: "Three months maybe, six months definitely."

March 2018: Self-driving will "encompass essentially all modes of driving" by the end of 2019 and be "100% to 200% safer than a person."

February 2019: "Feature complete" by end of year, fully confident drivers will not need to touch the wheel by Q2 2020.

January 2020: Tesla's self-driving feature would be so capable that drivers could "snooze in the driver seat" by the end of 2020.

January 2022: "I would be shocked if we do not achieve Full Self-Driving safer than human this year."

November 2025: Unsupervised FSD promised for select US cities within months. Musk claimed drivers would be able to text while driving "in a month or two."

January 2026: After failing to deliver unsupervised FSD by the end of 2025, Musk stated Tesla needed 10 billion miles of data. That number was itself a moved goalpost: he had previously indicated 6 billion miles would be sufficient.

April 2026: During Tesla's Q1 2026 earnings call, Musk confirmed unsupervised FSD for consumer vehicles would not arrive until Q4 2026 "at the earliest." Observers noted this was the tenth consecutive year Musk had made a similar promise.

There is also the hardware betrayal. Tesla officially confirmed in April 2026 that vehicles equipped with Hardware 3 (HW3) will never be capable of unsupervised FSD. Some HW3 owners paid between $8,000 and $15,000 for the FSD package on the explicit understanding that the hardware was sufficient for full autonomy. Tesla is now offering a discounted trade-in program instead of the free hardware upgrade it previously promised. As we covered in Tesla: A Trillion-Dollar Bet on Promises, the valuation of Tesla (TSLA) has always incorporated a future that the data does not yet support.

The AGI Remix: Same Pattern, Different Product

Now apply the FSD pattern to artificial general intelligence.

April 2024: Musk tells an X Spaces audience that AI could surpass the best human by 2025 or 2026.

Late 2025: Musk confidently states AGI will arrive in 2025. With weeks remaining in the year, he revises to 2026.

January 2026: In a 173-minute podcast, Musk presents a "clear timeline: achieve AGI in 2026. Within three years, robots will surpass surgeons. By 2030, the collective intelligence of AI will exceed that of all humanity."

Davos 2026: Musk sharpens the window to "by year-end."

The structural similarity to FSD is impossible to miss. A bold prediction. A deadline. A missed deadline. A revised deadline with adjusted language. A new prediction that is always, without exception, "next year or the year after." And underneath each revision, a fundraising round or a valuation event that the prediction directly supports.

Musk recently told xAI employees that the company is raising between $20 billion and $30 billion in funding per year. Talk of imminent AGI is not incidental to that fundraising. It is the pitch. As we argued in AI: Knowledgeable but Not Intelligent, the reality of AI development is that these systems are deeply knowledgeable but not genuinely intelligent, and the gap between the two is not a software update. It is a paradigm.

The definition of AGI itself is a moving target. Musk's definition, "smarter than the smartest human," is conveniently vague enough to be declared achieved at any moment that serves a business need, or deferred indefinitely when scrutiny arrives.

The Cost of Speed

The "build fast" philosophy has a cost that extends beyond GPU utilization and financial losses. It lands on real communities.

To power its Colossus data centers, xAI installed dozens of methane gas turbines in South Memphis and Southaven, Mississippi, without obtaining the required Clean Air Act permits. As of May 2026, the NAACP has sued xAI for operating 33 unpermitted gas turbines at its Southaven facility. The facility sits approximately half a mile from homes and one mile from an elementary school.

The emission numbers are staggering. According to Earthjustice, the unpermitted turbines have the potential to release up to 2,508 tons of smog-forming nitrogen oxides per year, making xAI likely the largest industrial source of NOx in Memphis. They could also emit up to 236 tons of fine particulate matter, 500 tons of carbon monoxide, and 19 tons of formaldehyde (a carcinogen) annually. Rather than address the violations, xAI added six more unpermitted turbines.

The community context makes this worse, not better. Memphis was recently named an "asthma capital." Both Shelby County, Tennessee, and DeSoto County, Mississippi, received an "F" for ozone pollution from the American Lung Association. The Boxtown neighborhood, closest to the xAI facility, has a cancer risk four times the national average. These are predominantly Black communities with decades of industrial pollution exposure. A new study commissioned by the Southern Environmental Law Center found that the proposed permanent turbines would impose an estimated $30 to $44 million in annual health damages.

Building the world's largest supercomputer in 122 days is an impressive engineering timeline. Doing so by running unpermitted gas turbines next to an elementary school in a community already failing federal air quality standards is not innovation. It is externalizing costs onto people who do not have the resources to fight back.

The Bull Case: What the Data Could Say Tomorrow

Credibility requires presenting the strongest version of the opposing argument. Here it is.

SpaceX is, by almost any measure, one of the most impressive engineering companies ever built. It has fundamentally transformed the commercial launch industry. Starlink has over 9 million users worldwide and generates real, growing revenue. The Falcon 9 is the most reliable and frequently launched rocket in the world. These are not promises. These are operational realities. SpaceX's reported $16 billion in 2025 revenue is backed by actual contracts with NASA, the Department of Defense, and commercial customers. Anyone dismissing the entire Musk enterprise as vaporware is ignoring the rocket that keeps launching.

The Colossus hardware exists. It is real. 550,000 GPUs are installed and running. The utilization problem is a software problem, and software problems are, in principle, solvable. xAI has set a target of 50% utilization. If they achieve even half of that improvement, the economics change materially. The IPO could raise between $40 billion and $80 billion in capital, which would fund the kind of software infrastructure buildout needed to close the utilization gap.

Grok usage is growing. The chatbot had roughly 64 million monthly active users as of early 2026, and the Grok app has been downloaded nearly 100 million times. Grok reportedly reached third place among US chatbots behind ChatGPT and Gemini. Revenue is growing, even if it remains small relative to losses.

And the Cursor deal, while structured unusually, gives SpaceX an option on one of the fastest-growing companies in AI coding. If the acquisition closes at $60 billion, SpaceX would own a product competing in the hottest category in enterprise software. The combination of Cursor's distribution with Colossus's hardware could create something neither company has alone.

The bull case is real. The question is whether it justifies a $1.75 trillion IPO valuation for a combined entity doing approximately $20 billion in revenue.

The Bottom Line

There is a pattern here. It is not subtle. It operates in a predictable cycle: a bold public prediction ("AGI by 2025," "full self-driving in two years," "a million robotaxis by 2020"). A deadline. A fundraising or valuation event that the prediction supports. A missed deadline. A revised prediction, with adjusted language and a new deadline, typically 12 to 18 months out. And then the cycle repeats. It has repeated on FSD for ten consecutive years. It is now repeating on AGI.

The Colossus lease to Anthropic is the most recent data point in this pattern. Musk built what he said was the world's fastest supercomputer to achieve AGI. Then his entire founding team left. His internal leadership called the utilization "embarrassingly low." His company burned through $7.8 billion in nine months. And now he is leasing the flagship facility to a direct competitor while preparing the largest IPO in history.

None of this means Musk is incapable of building successful companies. SpaceX is proof that he can. But it does mean that investors evaluating xAI, SpaceX, Tesla (TSLA), or the upcoming IPO should apply the same scrutiny to Musk's timelines that they would apply to any company's forward guidance. The track record on delivery is public. It is documented. And it is consistent.

At Wealth Engine Pro, we evaluate what is, not what someone says will be. The data on xAI is not ambiguous. $250 billion in valuation. 11% GPU utilization. $1.46 billion in quarterly losses. $107 million in quarterly revenue. 11 of 11 cofounders departed. 80+ total departures. 33 unpermitted gas turbines. And a decade of deadlines that arrive right on time to support a fundraise, but never quite arrive as a product. The numbers do not have a narrative. They do not need one.