The Bet

I want to tell you two things. One is idealistic. The other is a spreadsheet. They’re connected.

The idealistic thing: Right now, at this exact moment, the tools available to an independent builder are better, cheaper, and more powerful than anything a well-funded startup could access three years ago. A $1,500 laptop runs production-grade language models. DeepSeek’s API costs 90-95% less than GPT-5. Open-source models ship under MIT licenses. A solo developer with taste, a clear problem, and a weekend can build what used to require a team and a seed round.

The gates are open. That part is real.

The spreadsheet thing: The companies providing these tools are spending $602 billion this year on AI infrastructure — data centers, GPUs, cooling systems, power — to earn roughly $60 billion in traceable AI revenue.

That’s a 10-to-1 ratio. For every dollar coming in, ten are going out.

And the companies bridging that gap? They’re borrowing. They’re selling ads. They’re raising money from investors who expect the gap to close — eventually.

Both of these things are true at the same time. Your tools have never been this good. The economics funding them have never been this fragile.

This is the $602 billion bet.

(A note: Issue #2 teased “$660 billion.” As 2026 projections firmed up, CreditSights revised the combined hyperscaler capex figure to $602 billion. The economics got slightly less dramatic. The ratio didn’t.)

Let’s make the invisible visible — but let’s focus on the numbers that actually change what you build.

In 2026, the five largest cloud and AI companies — Amazon, Alphabet, Meta, Microsoft, and Oracle — will spend a combined $602 billion on capital expenditure, up 36% from last year. About 75% of that, roughly $450 billion, goes directly to AI infrastructure. US tech capex has reached 1.9% of GDP — exceeding the Apollo program, the Interstate Highway System, and the Manhattan Project combined. Goldman Sachs projects $1.15 trillion in total hyperscaler capex between 2025 and 2027.

These companies are consuming 90% of their operating cash flow on this buildout. They’re borrowing over $400 billion more to cover the gap.

Those numbers are staggering. But they’re not the ones that affect your daily work. These are:

The two companies diverging

OpenAI and Anthropic are the providers most independent builders interact with directly. And they’re heading in different directions.

OpenAI crossed $20 billion in annualized revenue by end of 2025. Against that: $14 billion in projected losses for 2026. The company spends $1.69 for every $1 it earns. HSBC projects it likely won’t turn a profit by 2030 and faces a $207 billion funding shortfall. Cumulative losses through 2028 are projected at $44 billion. It’s currently raising up to $100 billion at a valuation of $730-830 billion — roughly 40 times revenue.

ChatGPT still has 800 million weekly users. But its share of AI web traffic fell from 86.7% to 64.5% between January 2025 and January 2026. And on February 9, 2026, ChatGPT launched ads — sponsored products at the bottom of conversations on its Free and Go tiers. Minimum advertiser commitment: $200,000. Pro, Business, and Enterprise remain ad-free.

Anthropic’s trajectory diverges. The company reached $19 billion in annualized revenue by March 2026 — up from $1 billion fifteen months earlier. Eighty percent of that comes from enterprise customers. The burn rate is dropping: 33% of revenue in 2026, projected at 9% by 2027, with positive cash flow expected by 2028. Epoch AI projects Anthropic could surpass OpenAI in total revenue by mid-2026. Anthropic has pledged to remain ad-free.

This isn’t a company comparison. It’s a preview of two possible futures for how your tools get funded. One path leads through advertising — your conversations become inventory. The other path leads through enterprise value — your productivity becomes the product. The provider you build on is a bet on which future you think wins.

The circular loop

There’s one more thing worth seeing.

When OpenAI raises money from Microsoft, a significant portion flows back to Microsoft for Azure compute. When SoftBank invested $40 billion in OpenAI’s Stargate project, the proceeds flow to Stargate’s corporate partners. Nvidia committed up to $100 billion to OpenAI — and OpenAI’s CFO acknowledged that money goes right back to Nvidia for GPUs. Nvidia is also a prominent investor in CoreWeave, which supplies cloud infrastructure to OpenAI using Nvidia chips.

This is not inherently fraudulent. But it is circular. The same dollars are being counted multiple times — as investment, as revenue, as proof of demand — as they loop through the system. Hyperscalers raised $108 billion in debt in 2025 alone, with $1.5 trillion projected over the coming years.

The question isn’t whether AI is valuable. It’s whether $602 billion worth of infrastructure can find $602 billion worth of customers before the debt comes due.

Numbers describe the landscape. People navigate it. And the builders worth watching right now are the ones who’ve already made the choices that the rest of us are only starting to face.

Pieter Levels earns over $300,000 a month running a portfolio of web products — NomadList, RemoteOK, PhotoAI, and dozens more. Zero employees. PHP and SQLite. The story that usually gets told about Pieter is the romantic one: solo founder, laptop lifestyle, building from anywhere.

The story that matters more is the one about his 70 failures. Every dead project taught him something about dependency. His stack is simple not because he’s idealistic about simplicity — it’s simple because simple is cheap to switch. When an AI provider raises prices, he migrates before lunch. When a new model drops that’s 90% cheaper and good enough, he’s already running it by the afternoon. He doesn’t optimize for the best tool. He optimizes for the lowest switching cost.

That sounds like a technical preference. It’s actually a survival strategy. When the tools you depend on are funded by companies burning $14 billion a year, the ability to leave is the most valuable feature you can design for.

MiniPay tells a different kind of story. Built on the Celo blockchain in Africa, it has 12.6 million wallets and has processed over 350 million transactions. In Nigeria — where $92.1 billion in crypto value was received last year and 36% of the population remains unbanked — MiniPay isn’t a crypto product. It’s how people pay for things.

Most of its users don’t know they’re using blockchain. That’s not a failure of education. It’s a design philosophy: the infrastructure should be invisible; the utility should not. MiniPay didn’t wait for the infrastructure question to be settled. It built something useful on top of whatever infrastructure existed, made the complexity disappear, and let the value speak.

DeepSeek shifted the cost conversation entirely. A Chinese lab trained a model competitive with GPT-4 for $6 million instead of $100 million or more. V3.2 runs at $0.28 per million tokens — versus $1.25 for GPT-5. Open-source under the MIT License. The privacy trade-off is real — data routes through China, so it’s not suited for sensitive work. But for the vast majority of builder use cases, the quality is on par with Western models at a fraction of the price. V4 is announced for mid-2026.

DeepSeek matters not because it’s the best model. It matters because it proved that efficiency can compete with scale — and that proof gives every builder an exit door if their current provider’s economics change.

Each of these stories shares a thread: the builder’s survival doesn’t depend on any single company’s business model staying the same.

Here’s the non-obvious thing about the $602 billion bet.

Most bubble narratives have a clean ending. The music stops. The overextended companies collapse. The survivors pick up the pieces. The 2000 dot-com crash. The 2008 financial crisis. A sharp break, then a reset.

This one won’t work like that — and understanding why changes how you should build.

The circular financing loop means the system is deeply interlocked. Nvidia funds OpenAI. OpenAI buys Nvidia’s chips. Microsoft funds OpenAI. OpenAI runs on Azure. SoftBank funds Stargate. Stargate buys from SoftBank’s portfolio companies. Everyone’s revenue is partially everyone else’s investment. When the adjustment comes — and adjustments always come — it won’t be a single pop. It will be a slow, uneven deflation. Some subsidies will disappear quietly while others persist. Prices will shift on different timelines for different providers. The free tier that vanished wasn’t a market signal — it was one company’s CFO looking at one quarter’s burn rate.

This means the builders who navigate it successfully won’t be the ones who predicted the crash. They’ll be the ones who noticed which specific subsidy disappeared first — and had already built the muscle to adapt.

The historical parallel that fits best isn’t the dot-com bust. It’s the 1990s fiber-optic buildout. Companies like Global Crossing and WorldCom spent billions laying cable, went bankrupt, restructured or dissolved — and the cable stayed in the ground. The infrastructure survived. The internet we use today runs on fiber laid by companies that no longer exist. The value was real. The business models weren’t.

Azeem Azhar notes that the dot-com telecom buildout had a 4-to-1 spending-to-revenue gap. The railroad bubble of the 1870s: 2-to-1. This AI buildout is running at 10-to-1. The ratio is historically unprecedented. But the pattern — overinvestment creating infrastructure that outlasts the investors — is not.

So the $602 billion will likely build something genuinely valuable. And the companies spending it will likely face a reckoning. Both things. Same time.

What this actually means for your choices:

The Anthropic/OpenAI divergence isn’t gossip — it’s a structural fork. One company is moving toward ad-funded AI where your conversations are inventory. The other is moving toward enterprise-funded AI where your output is the product. When you pick a provider, you’re not just choosing a model. You’re choosing which funding pressure will eventually reshape your tools. Think about which pressures you can live with.

The circular loop means price signals will be noisy. A price drop might mean genuine efficiency gains — or it might mean one company is subsidizing your usage with another company’s investment. You can’t always tell the difference in real time. What you can do is ensure your product works at multiple price points. Test it at 2x your current API cost. If it breaks, that’s not a future problem. That’s a current vulnerability.

And the floor is rising. A $1,500 laptop runs production-grade models locally. DeepSeek offers frontier-competitive inference at 95% less. The distance between “cloud-dependent” and “self-sufficient” is shrinking every quarter. The builders who treat local capability as a hedge — not a hobby — will have options that others won’t when the subsidies shift.

Build something that works regardless of who wins the bet.

Other people are thinking about this clearly. You should read them.

Derek Thompson wrote “This Is How the AI Bubble Will Pop” — the clearest articulation I’ve found of how AI can be genuinely transformative and a financial bubble simultaneously. His comparison of AI infrastructure spending to national GDPs makes the scale visceral. Read it here

Ed Zitron at Where’s Your Ed At has been doing forensic work on OpenAI’s financial disclosures versus leaked internal numbers, finding billions in unexplained cash burn. Adversarial, detailed, and a necessary counterweight to the optimism cycle. Read it here

Epoch AI published “Anthropic Could Surpass OpenAI in Annualized Revenue by Mid-2026” — a rigorous, data-driven comparison of two companies on diverging trajectories. Free and Creative Commons licensed. Read it here

Lenny Mendonca and Martin Neil Baily wrote “An AI Bubble Won’t Trigger a Financial Crisis” in Project Syndicate — arguing that because hyperscalers are self-funding the buildout, the systemic risk is lower even if individual companies fail. An important counterpoint to the doom narrative. Read it here

The tools have never been this powerful. A solo builder with taste, determination, and a clear problem to solve has more leverage than at any point in history.

The subsidies funding those tools have never been this precarious. Six hundred billion dollars is looking for a return. When it starts finding one — or stops looking — the terms will change.

Build something that survives either outcome.

Next week: purpose-built silicon, local AI, and what happens when the chip can only do one thing — really, really well.

Go Beyond is a weekly newsletter tracking where the world is opening up — and where it’s closing.

If this issue made something visible that wasn’t before, share it with someone who’s building.