The Reckoning Behind the Revenue

When the Numbers Don't Add Up

AI is in the red.

That’s not surprising. Building the future has never been inexpensive. What surprises me is not that people aren’t asking about a path to profitability. They are. It’s that the assumption seems to be that if enough money is invested, enough people adopt, and enough capabilities emerge, the economics will eventually sort themselves out.

Maybe they will.

But assumptions are not business models. And when I started putting the numbers in the same room, I found a story that looks very different from the one most people are telling. The story isn’t that AI is expensive. The story is that the risk of all that spending is quietly moving from the people who created it to the people least equipped to see it coming. The losses are not the headline. Who ends up holding them is.

There is a number that should be sitting at the center of every boardroom conversation about AI right now. It isn’t the productivity gain. It isn’t the efficiency multiple. It isn’t the competitive moat being built or the talent being freed up for higher-value work.

It’s this: OpenAI, the company that started this revolution, projects $14 billion in losses in 2026 alone. Cumulative losses between 2023 and 2028 are expected to reach $44 billion.^[i] It also has 900 million weekly users and $20 billion in annualized revenue. A path to profitability exists on paper. It just doesn’t arrive until 2030, and it depends on several things going right in a market where nothing stays fixed long enough to model with confidence.

I’m not sharing this to suggest the technology doesn’t work or that the companies building it aren’t extraordinary. They are. The foundational work being done at OpenAI, Anthropic, Google, and others is remarkable, and what has been built in the last decade is what makes the next chapter possible. We stand on those shoulders. What we’re building wouldn’t exist without them.

And precisely because we are building in this space, I think the economics deserve an examination that most conversations are still avoiding.

Who Is Actually Carrying the Risk?

Here is what happens when you put the numbers together.

The AI providers are losing money despite massive and growing revenue. The enterprises buying their services are overpaying for something with no clean ROI framework. And here is the detail that almost never makes it into the same conversation: the prices enterprises are paying today don’t reflect the true cost of inference. The providers are absorbing a significant portion of that cost inside their losses. Enterprises are being onboarded at subsidized rates that the market hasn’t fully priced yet. When these companies eventually reach the profitability they’re projecting, through architectural improvements, pricing adjustments, or both, enterprise bills will almost certainly increase. The current pricing is not stable long-term pricing.

Everyone is spending more. The true cost of what they’re buying is higher than what they’re being charged. And the returns aren’t materializing at the rate the business cases promised.

This is not a crisis. It is a redistribution of risk. And anyone who has watched technology markets long enough will recognize it. Every market has uncertainty. The question is not whether uncertainty exists, the question is who is carrying it.

Right now, AI providers are carrying it through losses. Investors are carrying it through increasingly ambitious valuations. Enterprises are carrying it through spending commitments made before clear ROI frameworks exist. Employees are carrying it through workforce reductions justified by outcomes that have not always materialized.

The risk has not disappeared. It has simply been distributed across the system.

And the redistribution is not finished. Both OpenAI and Anthropic are moving toward public markets, with OpenAI already valued at $852 billion in private markets and Anthropic in early IPO discussions potentially targeting late 2026. In practical terms, this means the next wave of AI infrastructure may be financed by public market investors betting on profitability that the companies themselves don’t expect for three to four years. That is not unprecedented in technology. But it does mean the risk currently carried by providers, institutional investors, and enterprises is about to find a new home. Public market participants will be the next layer in this redistribution, and they deserve to understand what they are being asked to fund.

We Have Been Here Before

In the late 1990s, AOL and Yahoo defined the internet. They had the users, the revenue, the brand recognition, and the infrastructure. What neither of them had was a fundamentally better way to solve the core problem.

Google didn’t win because it had more resources. It won because it approached the problem from a different starting point entirely. PageRank treated hyperlinks as votes of authority rather than simply indexing keywords, and the results were noticeably better at a level ordinary users could feel immediately. That architectural difference, the decision to measure relevance differently at the foundation, made everything that came before look like a prototype. Yahoo’s portal strategy, AOL’s distribution dominance, none of it could compensate for building on a foundation that was less accurate at the thing that mattered most.

BlackBerry had enterprise email locked up. The security, the keyboard, the relationships with IT departments. Apple was a computer company. Until it wasn’t.

The pattern is consistent: early dominance does not protect against displacement when a competitor approaches the core problem from a fundamentally different foundation. What the incumbents have, distribution, enterprise relationships, capital, and brand, buys time. It does not buy permanence.

We are early. The infrastructure of the next chapter of AI is being built right now. Not in the gleaming campuses of the incumbents, but in the places where builders who have seen the structural limits clearly enough to approach the problem from a completely different starting point are doing the work. In garages. In university labs. And in our case, in a basement data center we call Firefly.

Most of what will take this industry forward has not been built yet. That is not a warning. That is the opportunity.

The question for anyone paying attention, whether investor, enterprise leader, or builder, is not whether AI has a future. It is which architecture that future runs on. And the companies currently dominating may not be the ones that define it.

What the Numbers Are Actually Saying

The provider losses are structural, not temporary growing pains. The reason is the same one we explored in the last piece: the transformer model was designed for research, not for the economics of running billions of conversations daily at civilization scale. The architecture creates costs that revenue, however impressive, cannot currently outrun.

Anthropic’s trajectory is meaningfully different from OpenAI’s and worth understanding. The company passed OpenAI in annualized revenue in April 2026, reaching $30 billion against OpenAI’s $24 to 25 billion, having grown 30 times in 15 months.^[ii,iii]

Anthropic projects positive cash flow by 2027, three years ahead of OpenAI. The divergence reflects a fundamental strategic difference: Anthropic’s enterprise-first focus generates more revenue per dollar of training spend. That is a meaningful signal about where the economics of this industry are heading.

But even Anthropic is still losing money. The entire frontier AI industry is, by design, betting that future revenue will justify present losses.

The path to profitability they’re projecting relies on several things happening simultaneously: revenue scaling faster than compute costs, inference getting dramatically cheaper through next-generation hardware, and agentic workflows generating higher revenue per user as AI moves from chat to complex multi-step tasks. It’s a plausible path. Amazon and Netflix took similar roads. But it requires multiple things going right in a market moving faster than any single company can fully control.

The 2029 and 2030 profitability timelines also carry weight for the capital markets question already noted above. As those public offerings approach, the assumptions underlying the valuations, about token costs, competitive dynamics, enterprise adoption curves, and architectural stability, were modeled in a landscape that may look unrecognizable by 2027.^[iv]

In practical terms, the next phase of the AI buildout may be financed by public market investors betting on profitability timelines that the companies themselves acknowledge are three to four years out. That is not unprecedented in technology. Amazon and Netflix lost money for years before the model worked. But those bets were placed in markets where a year felt like a year. In the current AI landscape, a month can feel like a year. That is not cynicism about the technology, but rather an accounting of the nature of the bet being placed. And the people being asked to fund it deserve to understand that clearly.

On the enterprise side, the numbers tell their own story. 92% of enterprises plan to increase AI spending over the next three years. Only 1% consider their AI strategies mature.^[v]

Average enterprise AI spend is projected to reach $11.6 million in 2026, a 65% jump from 2025,^[vi] yet only 29% of organizations report significant ROI from generative AI. Forrester already projects enterprises will defer 25% of planned 2026 AI spend into 2027 as the bills arrive ahead of the returns.^[vii]

Before going further, one thing needs to be clear, because it is the most misread part of this entire story. I am not arguing that AI costs will stay high or that they won’t fall. They are falling, fast, and they will keep falling. The point is stranger than that. Unit costs are dropping and total risk is rising at the same time. Both are true. Understanding why is the whole game.

This is where it gets counterintuitive. Let’s look at the dramatic reduction in token prices. The average cost per million tokens across major providers dropped from roughly $10 to $2.50 in a single year. Epoch AI research suggests inference costs are declining at rates approaching 200 times per year when accounting for both pricing and efficiency improvements. Andreessen Horowitz has coined the term LLMflation to describe this deflationary curve, drawing a parallel to Moore’s Law in semiconductors.^[viii]

NVIDIA’s next generation Rubin platform targets a ten-times reduction in inference costs compared to its current architecture.^[ix]

And yet enterprise AI bills are going up.

The reason is behavioral. When tokens get cheaper, teams run larger prompts, longer agent loops, more tool calls, more retries. A 99.7% token price decline has not reduced total AI spend, because cheaper calls encouraged much larger workflows, and most production costs moved outside the model invoice entirely into orchestration, vector databases, data egress, observability, compliance, and engineering time that never appears on a token bill.^[x]

Sam Altman acknowledged this directly in May 2026, saying the spending problem had become “kind of a meme now” among enterprise clients who had burned through their entire annual AI budgets in the first quarter.^[xi]

The frontier of this paradox sharpened further in June 2026. Anthropic launched Claude Fable 5 on June 9, its most capable publicly available model. Priced at $10 per million input tokens and $50 per million output tokens, double the previous top tier, it carries a one million token context window and up to 128,000 output tokens per request. The early feedback was exceptional. Independent evaluators reported breakthrough performance on complex, long-running tasks. The capability was clearly there.

What the early reviews didn’t capture, because there wasn’t time, is what Fable-class capability costs at scale. At that context window and output capacity, a single complex agentic task doesn’t just cost more per token. It consumes tokens at a scale that makes the LLMflation savings functionally irrelevant. When enterprise teams deploy frontier models on agentic workflows, the bill doesn’t increase linearly. The combination of larger contexts, extended reasoning chains, multi-step tool use, and retry loops means costs compound in ways that no quarterly budget modeled in advance. At frontier capability levels, you generate a great many tokens before you find out whether the task was worth it.

Fable 5 was pulled from general availability on June 12, three days after launch, following a U.S. government export directive. The directive was geopolitical in nature, not a response to performance, safety concerns, or anything the enterprise community did or could have influenced. The question of who controls access to frontier AI capability, under what conditions, and with how much notice, is one this series will return to. What matters here is the trust dimension: the infrastructure enterprises are being asked to build on can be withdrawn overnight, for reasons entirely outside their control, with no committed timeline for return. The bill and the availability are both unpredictable. That is a new kind of risk, and it doesn’t have a line item in most enterprise AI budgets yet.

Tokens are getting cheaper. The bill is going up. The model can disappear. That is the economic reality most enterprises are discovering after the fact.

The experimentation era is over. The accountability era has begun. And the CFO is being asked to make a trillion-dollar category decision with a framework designed for software licensing.

The Layoff Equation That Isn’t Adding Up

There is a human dimension to this economic story that deserves its own examination. It is the same pattern as everything above, with a human face on it. The risk doesn’t vanish when a role does. It moves. And it tends to land on the people who had the least say in the decision.

Through the first quarter of 2026, approximately 20% of confirmed tech layoffs were explicitly linked to AI and automation by the companies themselves.^[xii]

That is a dramatic increase from 2025, when AI was cited as a factor in fewer than 8% of layoff announcements. The list is extensive and growing. Microsoft cut more than 15,000 workers while its CEO confirmed AI was writing 30% of the company’s code. Amazon cut 14,000 corporate roles citing AI-driven efficiencies. Klarna replaced 700 customer service workers with AI. Citigroup is targeting 20,000 reductions as automation handles middle-office functions.^[xiii]

Here is what the business case for these decisions consistently assumed: that the cost savings from displacement would materialize faster than the costs of transition. In many cases, that assumption is not holding.

A Gartner study published in May 2026 found that while 80% of companies that piloted AI or automation technology reported workforce reductions, the businesses cut jobs regardless of whether the technology was actually generating returns.^[xiv]

They cut first and measured later. In some cases, they are now measuring and finding the math does not work the way the business case promised.

The rehiring costs are real. The institutional knowledge lost is real. The time required to rebuild capability after a premature displacement is real. And there is a cost that almost never appears in an ROI calculation: the damage to organizational trust.

When employees watch colleagues displaced for a technology that doesn’t deliver on its promises, the invisible agreement that the system is worth showing up for begins to fracture. The people who survived the cuts are not relieved. They are watching. And what they are watching is whether the organization they work for makes decisions they can trust.

Most workforce reduction models calculate salary savings. Very few calculate trust loss. Yet trust may be the more expensive variable. When people believe decisions are being made thoughtfully, they will tolerate uncertainty. When they believe decisions are being driven by pressure, trends, or incomplete information, something far more difficult to rebuild begins to erode.

The balance sheet rarely captures that cost. That fracture doesn’t recover through a town hall or a memo from HR. It accumulates. It surfaces in engagement scores and retention numbers and the quality of work from people who are present in body but have mentally begun their exit. This is the third level of trust, organizational trust, being stress-tested in real time by decisions made without full information.

This is what it costs to move faster than the evidence supports.

Risk doesn’t disappear when jobs disappear. It changes hands. The organization may reduce payroll expense, but employees absorb uncertainty. Teams absorb disruption. Leaders absorb credibility risk. And when the expected gains fail to materialize, everyone discovers that the original business case transferred far more risk than it eliminated.

What Comes Next: Reckoning or Reassessment?

Will enterprises use this moment to ask different questions, or will they simply defer spending and return to the same approach when the pressure eases?

My take is that a bifurcation is already underway. Enterprises that found specific high-value use cases and built clear measurement frameworks around them are doubling down. Enterprises that deployed broadly without a strategy are quietly pulling back, trying to figure out what they actually bought. The reckoning isn’t one dramatic moment. It’s already happening incrementally, in budget reviews and board conversations where someone finally asks for the ROI evidence and the room goes quiet.

The ones who use this moment to ask different questions, about architecture, about what AI is actually for, about what responsible deployment looks like, will be better positioned for what comes next. The ones who wait for the technology to improve enough to paper over the current gaps will find themselves in the same conversation again in three years.

This is not a break-the-glass moment. It is an inflection point. And inflection points are where the next chapter gets written.

The companies building that next chapter aren’t optimizing the transformer. They aren’t wrapping existing models in a new interface. They are building relational intelligence infrastructure designed from the ground up for the human and organizational realities that current AI consistently fails to address. Infrastructure that understands that the most expensive token is the one that didn’t need to be generated. That the most valuable AI interaction is the one that builds trust rather than eroding it. That the measure of a successful deployment is not efficiency gained but capability grown, in the people and organizations the technology serves.

That category doesn’t have a widely accepted name yet. That is usually a sign you’re early to something important.

We don’t have to ask if AI is creating value. It is.

What remains less certain is how that value will ultimately be distributed, what risks it depends upon, and who absorbs the consequences when expectations outrun reality.

Every technology wave creates winners. The question is whether it also creates understanding. Right now, we are measuring adoption, investment, valuation, and spending. What I am less certain we are measuring is who is left holding the risk when the assumptions underneath those numbers stop holding.

That is the real reckoning. Not whether AI works. Whether we are asking enough questions before we build the future on top of it.

In the next piece, we look at what happens to the founders and enterprises building on top of the current AI platforms when those platforms consolidate. Because the ground beneath the current buildout is less stable than most people realize.

Trust isn’t built on promises. It’s built on understanding the tradeoffs before the bill arrives.

Because when the returns come, everyone celebrates.

When they don’t, someone still pays.

This is Part 2 of The AI Reckoning Series: The Reckoning Behind the Revenue, When the Numbers Don’t Add up.

Read Part 1: The Hidden Cost of Intelligence, The Trust Story Hiding in Plain Sight

Endnotes

ⁱ https://europeanbusinessmagazine.com/sam-altmans-openai-is-burning-billions-most-users-pay-nothing-as-anthropic-closes-in/

ⁱⁱ https://www.the-ai-corner.com/p/anthropic-30b-arr-passed-openai-revenue-2026

ⁱⁱⁱ https://www.saastr.com/anthropic-just-passed-openai-in-revenue-while-spending-4x-less-to-train-their-models/

^iv https://tech-insider.org/anthropic-vs-openai-2026/

^v https://onereach.ai/blog/what-shapes-enterprise-ai-agents-in-the-future/

^vi https://writer.com/blog/enterprise-ai-adoption-2026/

^vii https://bizzdesign.com/blog/enterprise-ai-adoption-balancing-innovation-and-roi-2026

^viii https://www.artefact.com/blog/is-ai-really-getting-cheaper-the-token-cost-illusion/

^ix https://www.investing.com/analysis/the-ai-token-pricing-crisis-behind-openai-and-anthropics-revenue-race-200680777

^x https://www.navyaai.com/reports/ai-cost-report-token-prices-vs-ai-bill

^xi https://www.tomshardware.com/tech-industry/artificial-intelligence/openai-ceo-sam-altman-admits-ai-token-costs-are-becoming-a-huge-issue

^xii https://tech-insider.org/tech-layoffs-2026-ai-workforce-impact/

^xiii https://tech.co/news/companies-replace-workers-with-ai

^xiv https://fortune.com/2026/05/11/ai-automation-layoffs-gartner-study-roi/