The OpenAI Math Problem: Two Sides of the Most Expensive Bet in Tech History

Bull case. Bear case. Both are real. Here's the framework business leaders should use to place their own bets.

Every business leader betting their AI roadmap on OpenAI is implicitly placing one of two trades — the bull's "next AWS" or the bear's "next WeWork" — and most haven't done the work to know which one. The capital story makes the headlines. The unit economics tell a different story beneath the surface. The competitive picture has shifted in the last sixty days in ways that change the math.

// OpenAI raised more capital than any private company in history — $122 billion at an $852 billion valuation — against $14 billion in projected 2026 losses and a $500 billion Stargate compute commitment that only partially pencils out.

// Anthropic just passed OpenAI on absolute revenue while spending four times less on training, per SaaStr's analysis of the latest enterprise spend data — a signal that the frontier-model business is commoditizing faster than the bull case allows.

// OpenAI missed key pre-IPO milestones in early 2026 — including its 1B-weekly-user goal — surfacing tension between Sam Altman and CFO Sarah Friar over spending discipline and the timing of a public listing.

The leaders who win over the next two years won't be the ones with the strongest opinions about OpenAI. They'll be the ones who built their AI stack so that the answer to "what happens if OpenAI's economics break" is "we move workloads in a sprint, not a six-quarter migration."

I started this week annoyed at ChatGPT's interface. The new tab UI had moved twice in a month, my Memory configuration kept getting overwritten, and my custom GPTs were behaving like they'd been retrained against me. I wrote a grumpy LinkedIn post about it. By Thursday morning, I'd deleted the post.

The problem wasn't the interface. The problem was that I was looking at OpenAI through the wrong frame.

My first job in tech was at PSINet — one of the earliest commercial ISPs, in the middle of the original "this changes everything" infrastructure buildout. I had a front-row seat to the growth, the thrash, the over-investment, and, eventually, the fiber-optic bone yard the bust left behind. PSINet itself filed for bankruptcy in 2001 after spending billions on capacity that demand never quite caught up to. Global Crossing, WorldCom, McLeodUSA — same story, different logos. The technology was real. The economics weren't.

So I'll cop to it: when I read about $500 billion compute commitments and $852 billion private valuations, part of me is the old guy on the lawn yelling, "Get off my network." But there's a more honest reason I'm spending my Friday on OpenAI's income statement instead of writing about prompts.

The capital flowing into this round of AI buildout isn't just venture money from people who can absorb the loss. It's pension funds. It's university endowments. It's sovereign wealth. The same kinds of vehicles that lost real money on the 2000 telecom bust are funding the 2026 AI buildout — and a lot of them are past the point where "wait it out" is cost-free. That's why I don't want to see any of these companies fail. And it's also why I want to look at the numbers honestly, not optimistically.

So by Friday, I was studying OpenAI's reported income statement instead of complaining about Chrome. Here's what I found, and what I think it means for anyone running a business that depends — even indirectly — on whether OpenAI becomes a self-sustaining company by 2030.

The conventional take on OpenAI is some flavor of "they're winning, the numbers are eye-popping, the dependence on them is a fact of life." The numbers are real. The dependence is real. But "winning" obscures a more complicated picture: OpenAI is simultaneously the largest revenue-generating AI company on earth and the largest cash-burning private company on earth, in the middle of a competitive squeeze that wasn't supposed to arrive this quickly.

What follows is the bull case and the bear case, run honestly, scored against the actual evidence as of May 2026 — followed by a playbook for how a business leader should think about their own exposure.

One note before going further, the analysis below is OpenAI-specific because OpenAI is the largest, the most-watched, and the most-dependent-upon, but the dynamics generalize. Anthropic is burning cash chasing OpenAI. Google subsidizes Gemini against its search cash flows. Microsoft prices Copilot below cost to capture distribution. Perplexity, Mistral, Cohere, the open-weights labs — every frontier-model vendor in your stack faces some version of the same math. OpenAI is the canonical case. Read the playbook below as the template for anything you've built on top of any AI vendor.

The OpenAI math problem is the gap between three sets of numbers that all describe the same company.

The capital number is enormous. OpenAI closed a $122 billion round in March 2026 at an $852 billion post-money valuation — the largest private financing round in history. SoftBank led, with Amazon, NVIDIA, and a syndicate of hyperscalers and sovereign funds participating. Cumulative committed capital across equity, the Microsoft relationship, and the Stargate joint venture sits north of $200 billion.

The revenue number is impressive. OpenAI's annualized revenue hit $25 billion in February 2026, roughly double the prior year. ChatGPT counts 900 million weekly active users, the largest consumer software audience built at this growth rate.

The loss number is bigger than both, and that's the part that matters for any business leader treating OpenAI as a strategic vendor.

At the individual user level, the math problem looks like this: roughly 5% of ChatGPT's weekly users actually pay, meaning every query the other 95% submit costs OpenAI real compute money. Deutsche Bank's analysis (cited in the same WSJ piece) puts OpenAI's all-in spend at roughly $2.25 for every $1 of revenue. Sam Altman has publicly conceded that the $200/month Pro tier is unprofitable today.

// These numbers don’t work today

$2.25 of spend for every $1 of revenue.

The OpenAI math problem in a single ratio. Deutsche Bank's estimate, via the WSJ's April 2026 reporting.

At the enterprise customer level, the picture is different — but still not stable. ChatGPT Enterprise revenue is the fastest-growing line in the business. Once a 5,000-person company standardizes on it, switching costs become real. But the latest Ramp AI Index puts Anthropic at 34.4% of business adoption, compared with OpenAI's 32.3% in April 2026 — marking the first time Anthropic has surpassed OpenAI in business adoption. Among first-time AI buyers, Anthropic now wins about 70% of head-to-head matchups.

At the industry level, OpenAI is on track to incur $14 billion in losses in 2026 alone, with internal projections indicating $115 billion in cumulative losses through 2029. CFO Sarah Friar has reportedly warned the board that annual cash burn could climb to ~$57 billion by 2027 if growth doesn't accelerate. The capital cycle has to keep turning. Skip one round and the math changes everything.

The bull and bear cases agree on the inputs. They disagree on what compounds and what compresses.

The bull case unfolds in three stages.

Stage 1: AWS playbook, faster. Amazon spent ten years burning cash on infrastructure that became a profit machine. AWS now does $100B+ in revenue at roughly 30% operating margins. AWS hit $13B annualized revenue around year nine; OpenAI hit $25B in year three of commercial operations. The capex profile looks ugly until you remember what AWS looked like in 2008.

Stage 2: Inference cost collapse buys margin. Inference price per million tokens has dropped between 9x and 900x per year, depending on the task, with the median trend running around 200x annually since January 2024. Per-token API pricing has collapsed roughly 90% since early 2025. Most of OpenAI's "losses" are R&D and capex for capacity that comes online over a five-year window. The unit economics of a 2026 query running on 2028 inference costs look very different.

Stage 3: Distribution plus enterprise lock-in compounds. ChatGPT's 900M weekly users is a habit moat. Most have never tried Claude or Gemini and never will. Enterprise contracts, Memory configurations, custom GPTs, and integration tooling create real switching costs once a 5,000-person company is standardized.

The bear case unfolds in three stages too.

Stage 1: Telecom buildout, redux. In 1999, telcos laid fiber assuming demand would catch up to capacity. I watched this rollercoaster from inside PSINet — over 100 mergers and acquisitions rolled through the ISP industry in a few years, then the bust took most of them under. Global Crossing, WorldCom, McLeodUSA — the fiber became someone else's cheap asset. OpenAI is making the same bet on compute. Stargate is $500B of capex committed against an assumption that demand-per-dollar-of-capex will scale.

Stage 2: The technology is commoditizing. Frontier model quality is converging. Claude, Gemini, Qwen, DeepSeek — all within striking distance of GPT-5 on benchmarks that matter for business work. Anthropic just passed OpenAI on revenue while spending four times less on training. When technology commoditizes, margins compress; they don't expand. That's the opposite of the AWS analogy.

Stage 3: The capital cycle is the real product. OpenAI has to keep raising funds for the next training run to justify the next valuation to raise the next round. Fortune's April 28 coverage of the WSJ reporting captured the dynamic: missed pre-IPO milestones, ~$660 billion in 2026 AI capex commitments across the four major hyperscalers, and open friction between Sam Altman and CFO Sarah Friar over spending discipline. The capital cycle is a beautiful machine while it works and a catastrophe when it doesn't.

Four named mistakes leaders make when reasoning about OpenAI exposure.

Misstep 1: Assuming OpenAI is too big to fail. It is too big to disappear. I keep thinking about how Yahoo! was the leader in the 1990s and is now virtually non-existent. OpenAI is not too big to restructure, IPO at a markdown, or exit the consumer business to focus on enterprise. "Too big to fail" was also the working assumption about the 1999 telcos. The companies survived in name. The shareholders did not.

Misstep 2: Pricing strategic bets on current API rates. Per-token pricing has dropped 90% since early 2025 and will keep dropping. Pricing your business case on today's rates leaves margin on the table; pricing it on the assumption that today's rates persist leaves you exposed when they move. Use a band, not a point estimate.

Misstep 3: Treating model portability as a "later" problem. The migration cost from a stack tightly coupled to OpenAI to a model-agnostic stack is a function of how deeply OpenAI's quirks are baked into prompts, evals, and tooling. Every quarter you wait, the migration gets longer, not shorter. The competitors quietly building model-agnostic stacks now are buying themselves a 2027 advantage.

Misstep 4: Anchoring to revenue growth instead of margin progress. OpenAI's revenue chart is a flattering picture. The losses chart is the honest one. A vendor whose unit economics aren't yet positive is a vendor whose pricing has to move — either toward a sustainable model (more expensive for you) or toward a bankruptcy event (catastrophic for you). Track the margin signal, not the revenue signal.

// Key Takeaways

Don't make 5-year strategic bets on OpenAI's current pricing or current existence. Pricing will move. Existence is more probable than not, but not certain. If your AI roadmap has a single point of failure named OpenAI, fix that this quarter — not in 2027.

Build an AI strategy with model portability in mind today. Test critical workflows on Claude, Gemini, and at least one open-weights model. The competitors quietly making their stacks model-agnostic are buying themselves a 2027 advantage that compounds.

Watch the margin signal, not the revenue signal. OpenAI's $25B revenue is the flattering chart. The $14B losses on $2.25-of-spend-per-$1-of-revenue is the honest one. Vendor unit economics that aren't positive will resolve — toward sustainable pricing or toward a restructuring event. Either way, your costs move.

Treat the next OpenAI funding round as a board-level signal, not a press story. Flat rounds, inside-led rounds, extended timelines — those are the leading indicators that the bull thesis is breaking. They show up six to twelve months before the consequences hit your contract.

The temptation in 2026 is to pick a side. To go all-in on OpenAI because the consumer brand and enterprise distribution feel insurmountable. Or to bet against them because the unit economics look unsustainable. Both reactions are emotionally satisfying. Both are strategic mistakes.

The honest planning posture is: build for the bull case, hedge for the bear case, and re-evaluate quarterly. That means continuing to deploy OpenAI where it's the best tool — and most of the time it is — while making sure that the day you need to swap providers, the swap is a config change, not a six-quarter migration.

For the 2026 budget cycle, that translates to three line items: a model-portability audit (one engineering quarter), a model-agnostic gateway (off-the-shelf or in-house look at OpenRouter for the design pattern), and a quarterly OpenAI exposure review on the executive team's calendar. None of those line items is exotic. All of them are cheap insurance against the version of 2027 in which the bear case turned out to be early but right.

The most likely outcome isn't "OpenAI wins big" or "OpenAI collapses." It's something messier in the middle: a $50–80B revenue company at 10–15% operating margins by 2030, valued somewhere between $400B and $800B. Today's $852B price is either right or off by 2–3x. Those are very different outcomes for anyone holding equity, debt, or a long-term commercial contract.

Your AI Sherpa,