In 2019, OpenAI CEO Sam Altman admitted he wasn’t focused on monetizing the company, which had started as a nonprofit before becoming a capped-profit entity. “We have no current plans to make revenue,” he told investors. He went further: OpenAI didn’t plan on knowing when it might one day generate revenue. Instead, Altman offered what he called “a soft promise to investors that once we’ve built this sort of generally intelligent system, basically we will ask it to figure out a way to generate an investment return for you.” Since then, OpenAI has restructured into a public benefit corporation valued at $760 billion, dropped the word “safely” from its mission statement, and is now seeking a $100 billion funding round, in part to pay for infrastructure deals it has already committed to. Elon Musk, an early co-founder, sued to block the conversion; a federal judge denied his injunction but sent the case to trial.
Two years later, hyperscalers and frontier AI labs seem to be in a de facto race to outspend and outbuild their rivals. Only the best GPUs and biggest data centers will do. Only the best benchmarks. On Tuesday, Meta and AMD announced a chip deal reported to be worth up to $100 billion, one of the largest hardware procurement agreements in corporate history. It came a day after Goldman Sachs’ chief economist, Jan Hatzius, said the economic impact of AI investment on U.S. GDP growth in 2025 was “basically zero.” And it came two weeks after Alphabet issued a century bond, a debt instrument maturing in 2126, to help finance $185 billion in capital expenditures this year. The last technology company to issue a century bond was Motorola, in 1997.
If scaling solved the reliability problem, this market wouldn’t exist. But it does, and it’s growing fast.
Nvidia sells GPUs and an open-source guardrails toolkit to protect you from what the GPUs produce. Salesforce built “Agent Script,” a deterministic scripting layer that gives CIOs more responsibility in engineering and maintaining AI controls. The fix for the probabilistic AI agent was to stop it from being probabilistic.
❯ pre-commit hook running…
❯ validating concept_ids against schema
✗ pcic_12_1_cop_ser: NOT IN CANONICAL SCHEMA
✗ pcic_4_1: SECTION DOES NOT EXIST
✗ COMMIT BLOCKED — 46 invalid IDs detected# Cost: $0. Latency: 200ms. Accuracy: 100%.
# No model needed. No API call. No apology.
A Foundation Capital analysis put it: getting from 80 percent accuracy to 99 percent can take 100 times more work than the initial build. The demo works. Production doesn’t. The gap between them is where all these companies live.
The industry is spending hundreds of billions per year on model development. It is spending another $50 billion on the infrastructure to catch the model’s mistakes. One of the most effective tool in the stack, the one that actually stops bad output from shipping, is a deterministic gate. A binary check. Pass or fail. It doesn’t need a century bond. It doesn’t need a data center in orbit. It doesn’t need to go to eleven.
It just needs to say no.
The plan, in other words, hasn’t changed much since Altman’s confession. Build the infrastructure first. Figure out the business model later. If you build it, AGI will come. Trust us. But the zeros on the end have multiplied. In 2026, the four biggest U.S. technology companies are projected to spend between $635 billion and $665 billion on capital expenditures alone. Morgan Stanley estimates hyperscalers will borrow $400 billion this year to help pay for it. On the private side, the numbers are equally mind-boggling: OpenAI is reportedly fundraising at a valuation above $850 billion, Anthropic at $380 billion, and the newly merged SpaceX-xAI at $1.25 trillion. The three largest private funding rounds in tech history all closed within months of each other. And the definition of what all this spending is chasing, artificial general intelligence, remains, as Anthropic co-founder Daniela Amodei put it in January, “maybe not wrong, but just outdated.”
Meta is the latest entrant in a spending cycle that has consumed every major technology company simultaneously. Amazon has guided $200 billion in capital expenditures for 2026, more than the GDP of Greece, while cutting tens of thousands of jobs. Alphabet, which spent $52.5 billion on capex in 2024, has nearly quadrupled that figure to $175–185 billion, financing the gap with a $32 billion bond sale that drew nearly ten times its target in investor demand. Microsoft is on pace for $145 billion, even as it disclosed that 45 percent of its $625 billion cloud backlog is tied to a single customer: OpenAI. Collectively, the R&D expenditures of the five largest tech companies now exceed $250 billion annually, more than the federal government spends on all non-defense research combined. Their combined capex approaches $650 billion. And for the first time, their aggregate infrastructure spending, after buybacks and dividends, may exceed their projected cash flows. The companies that defined “asset-light” are now the most capital-intensive enterprises on Earth, financing data centers with century bonds, shell entities, and ay for itself.
The cross-pollination runs deeper than capex alone. In the past year, the major AI players have begun financing each other directly, investing in their own customers, taking equity in their own suppliers, and structuring deals that make it increasingly difficult to tell where one company’s balance sheet ends and another’s begins. Nvidia invested $5 billion in Intel, becoming a 4 percent stakeholder in its longtime rival. Nvidia has since walked away from an unfinished plan to invest up to $100 billion in OpenAI, pivoting to a roughly $30 billion equity investment as part of OpenAI’s current mega-round. Nvidia also committed up to $10 billion to Anthropic, both as an investor and as a compute supplier. AMD, fresh off an identical warrant-based deal with OpenAI in October 2025, just replicated the structure with Meta: up to 160 million shares at a penny apiece, vesting as Meta ramps purchases from one gigawatt to six. Nvidia also owns roughly 7 percent of CoreWeave, which has spent $7.5 billion buying Nvidia chips, meaning, as OpenAI CFO Sarah Friar acknowledged, “most of the money will go back to Nvidia.” Tesla invested $2 billion in xAI, which then merged with SpaceX in a deal valuing the combined company at $1.25 trillion, with Musk pitching space-based data centers as a way to meet AI’s energy demands. Anthropic, meanwhile, closed a $30 billion round at a $380 billion valuation, with participation from both Microsoft and Nvidia, companies to which Anthropic has committed $30 billion in compute purchases. Goldman Sachs flagged this pattern in June 2024, in a report titled “Gen AI: Too Much Spend, Too Little Benefit?” At the time, hyperscaler capex was roughly $250 billion. Goldman’s head of global equity research, Jim Covello, warned that AI technology “must solve complex problems to justify its costs, which it isn’t designed to do.” The industry’s response: triple the spend.
These go to eleven
In the 1984 mockumentary “This Is Spinal Tap,” guitarist Nigel Tufnel shows off an amplifier whose dials go to eleven. When the interviewer asks if that makes it louder, Tufnel replies: “Well, it’s one louder, isn’t it?” Asked why he doesn’t just make ten louder, Tufnel pauses, baffled. “These go to eleven.”
The AI infrastructure boom is a collection of gorgeous guitars hanging on a wall, each more expensive than the last, none of them yet plugged in. The AMD-Meta deal announced Tuesday? First shipments don’t start until the second half of 2026. Only the first gigawatt is a binding commitment; the other five are contingent on milestones. The $500 billion Stargate project? Still got the tag on it. Alphabet’s $185 billion capex guidance? Financed with a bond that matures when everyone in the room is dead. Meta’s Hyperion campus in Louisiana? Financed through a shell entity called “Beignet Investor LLC.” Don’t touch it. Don’t point even.
And the sustain? Goldman Sachs’ chief economist says the economic impact was basically zero in 2025. J.P. Morgan estimated last November that AI would need to generate over $600 billion in annual revenue just to earn a 10 percent return on the infrastructure investment.
When someone asks “why not just make ten louder?” why not make the existing models more efficient, fix the architecture, solve the memory problem, close the jagged frontier instead of building more data centers. If not just on earth, how about in space? Google launched a program to put AI chips in orbit. Elon Musk merged xAI with SpaceX in a $1.25 trillion deal explicitly to build orbital data centers, and filed with the FCC for authorization to launch up to a million satellites running inference in space.
Jagged intelligence, clipped signal
The term jagged intelligence was first described in a 2023 Harvard Business School field experiment with Boston Consulting Group consultants, where researchers found a “jagged technological frontier”: AI substantially improved performance on some realistic tasks but degraded performance on others that appeared similar. Andrej Karpathy, the former Tesla and OpenAI researcher, popularized the shorthand in 2024. In January 2026, economist Joshua Gans formalized it as “Artificial Jagged Intelligence” in a paper that proved mathematically what practitioners already knew: scaling improves the average level of reliability but does not change the shape of the heterogeneity. Put plainly, more compute raises the peaks without filling the valleys. Headline benchmarks improve. Surprising failures persist.
Sam Altman told investors years ago that the plan was to build the machine and then ask it how to make money. The machine is now financed with century bonds, circular equity swaps, shell entities named after pastries, and $400 billion in annual borrowing. It runs on GPUs that depreciate in three to five years but sit on balance sheets at six. It powers models that crush every benchmark and struggle to remember what you told them five minutes ago. It chases a definition of intelligence that the field’s own luminaries call outdated. And every six months, a new model drops. It scores higher. It costs more. It still can’t answer the question Altman promised it would. Sometimes it can’t even keep the year straight.
Well, it’s one louder, isn’t it?
