Last month, a client sent me a market brief with a single line that stopped me cold: enterprise AI spending would hit $177 billion by 2026, 'led by Anthropic, OpenAI, and Google.' She wanted to know if she should restructure her DAO's treasury allocation around it. I asked her one question — what counts as enterprise AI spending? She didn't know. The brief didn't say. And in that silence, I recognized a pattern I've watched repeat for nearly three decades in finance and six years in decentralized governance: a number so large it feels like gravity, attached to a claim so vague it can mean anything.
This matters because we are living through a period where institutional capital is migrating into crypto rails at unprecedented speed. Last year I sat across from BlackRock's venture arm negotiating a $10 million grant allocation. Their analysts quoted market-size projections the way some people quote scripture. If the numbers underpinning those negotiations are epistemically hollow — and this $177 billion figure is a textbook case — then the whole architecture of institutional adoption is being built on sand. Code without compassion is cold, but data without provenance is dangerous.
Let me be clear about what this figure is and isn't. The brief that circulated last month — appearing on a crypto media outlet, with no author, no methodology, no source institution, and a headline that repeated the summary verbatim — is not reporting. It's a data point laundered through publication. I've audited enough smart contracts to know that when someone hands you a number without a source, you treat it the way you'd treat a token with no audit: assume it's a honeypot until proven otherwise.
The sourcing vacuum isn't a minor detail. It's the entire analytical problem. The $177 billion figure only holds up under a broad definition — all enterprise AI spending, including infrastructure, professional services, internal builds, and traditional machine learning. Under a narrow definition — generative AI model APIs and AI-native SaaS subscriptions — the number is inflated by roughly seven to twelve times. Menlo Ventures reported mid-2025 that enterprise generative AI spending for the full year would land around $13.8 billion. Even adding private deployments and systems integration, you're looking at $20 to $30 billion. For $177 billion to be the narrow number, you'd need a six-to-twelve-fold year-over-year increase, which violates basic enterprise procurement physics: IT budgets are set annually, and a tripling is already extreme.
So the $177 billion is almost certainly a broad-definition figure. Fine. IDC's global AI spending path — $307 billion in 2025, $632 billion by 2028, including consumer — places enterprise-side 2026 spending comfortably in that range. But here's where the narrative breaks: a broad number cannot support a narrow ranking. The claim that Anthropic, OpenAI, and Google 'lead' enterprise AI spending only holds in the enterprise LLM API segment, which Menlo sized at roughly $8 to $10 billion in 2025 — Anthropic at 32%, OpenAI at 25%, Google at 20%. Under the broad definition, the firms capturing the most enterprise AI revenue are Microsoft, NVIDIA, AWS, and the systems integrators. The brief took a segment-level ranking and welded it onto a total-market figure. That's not analysis. That's a category error dressed as a forecast.
And I should note the irony: Menlo Ventures, the source for those API share numbers, is an Anthropic investor. I've spent enough time in rooms with investors to know that the data they commission tends to confirm the thesis they need confirmed. The directional finding — Anthropic leads enterprise LLM API spending, propelled by coding and agentic workloads — is plausible and matches what I see in governance tooling. But I discount the precision accordingly.
Here's what the $177 billion number actually tells us, and it's more uncomfortable than the headline suggests. This is not growth. It's reallocation. Enterprise IT budgets grow at 5 to 8 percent annually. They cannot support 6-to-12-fold AI spending increases. Which means the overwhelming majority of that $177 billion is being pulled out of existing software budgets — seat-based SaaS subscriptions migrating into consumption-priced intelligence. This is why we've watched traditional SaaS valuation multiples compress for two years straight. The AI spending boom is, for the software industry, a zero-sum shock dressed as an addressable market.
The value capture is even more lopsided than the narrative admits. OpenAI's annualized revenue sat around $13 billion at the end of 2025; Anthropic around $7 billion and climbing fast. Meanwhile NVIDIA's data center revenue alone exceeded $50 billion in a single quarter. Microsoft, Google Cloud, and AWS each report AI-related revenue dwarfing the model labs. 'Leading spending' and 'capturing spending' are entirely different claims, and the brief conflates them. If you want to know where the $177 billion actually flows, follow the GPUs and the cloud bills, not the API invoices.
What's systematically hidden from this narrative is equally instructive. The brief ignores China entirely — Alibaba's Qwen, DeepSeek, ByteDance's Doubao, and Zhipu are penetrating Chinese enterprise markets in ways that would materially reshape any truly global ranking. It ignores open-source model erosion: Llama, Qwen, and DeepSeek deployments inside enterprise firewalls don't appear in any API share statistic. And it ignores the conversion problem. MIT's NANDA research circulated a controversial finding that 95 percent of enterprise GenAI pilots produced no measurable P&L impact. Halve that number, and the linear extrapolation behind $177 billion collapses into speculation.
Then there's the deflation trap that no market-size forecast adequately handles. Inference prices are falling five to ten times annually. A dollar-denominated spending forecast is acutely sensitive to this: usage can grow tenfold while prices fall tenfold, leaving dollar revenue flat. If enterprises shift workloads toward smaller models, distilled models, or self-hosted open-source deployments — which is exactly where the economics push them — the model layer's revenue ceiling sits far below what $177 billion implies.
The contrarian read: the $177 billion figure's real function is not to guide procurement. It's TAM fuel for private-market valuations. Anthropic's reported $183 billion post-money valuation against roughly $7 billion in annualized revenue implies a 25x price-to-sales multiple. OpenAI's ~$500 billion valuation against ~$13 billion implies roughly 38x. Those multiples only survive under simultaneous assumptions of sustained multi-fold spending growth and uneroded share. The $177 billion number is the external validation for both assumptions. It is a narrative asset, not an analytical input.
What should governance architects and treasury managers actually do with this? First, demand three-line disclosures on any market figure: API spending, AI-native SaaS, and total including infrastructure. Refuse the single number. Second, track token volume and workload count rather than dollar figures — usage is real, revenue is a price-sensitive fiction. Third, recognize that the beneficiaries of enterprise AI spending are the picks-and-shovels layer, not the model layer, and position accordingly.
Based on my audit experience with governance treasuries, the institutions that survive the next cycle won't be the ones that quoted the biggest market number. They'll be the ones that asked what it measured. The $177 billion figure isn't wrong, exactly. It's just answering a question nobody in the room actually asked — and the gap between the question and the answer is where real capital gets destroyed.
We are building systems that will govern billions in collective resources. The least we can do is insist that the numbers we build them on can survive a single follow-up question. In a sideways market waiting for direction, the discipline of provenance is not a luxury. It's the position itself.