AI is slowing down…..
Ed Zitron’s “AI Is Slowing Down” hit Hacker News on Sunday and stayed pinned on the front page for most of the day. The argument is uncomfortable for the people building the industry, and the numbers behind it are uncomfortable too.
Zitron’s central claim: the AI industry has committed to $9.5–15 trillion in data-center buildout over the next six years, and that infrastructure demands at least $2 trillion in annual revenue by 2030 to make economic sense. Current trajectory falls short by an order of magnitude.
If you have been listening to AI keynotes for two years and feeling something does not quite add up — this is the post that names what.
What the post actually argues
Zitron, who runs the Where’s Your Ed At newsletter, has been the most-quoted AI skeptic of the cycle. His thesis here is sharper than his usual takes because it focuses on one specific equation:
- Hyperscalers and AI labs have planned roughly 190 gigawatts of new data-center capacity (Sightline Climate, February 2026).
- NVIDIA CEO Jensen Huang put the cost per gigawatt at $80–100 billion.
- Multiply those together and you get $9.5–15 trillion in committed infrastructure spend.
- The two companies driving virtually all that demand — OpenAI and Anthropic — project combined $358 billion in annual revenue by 2029.
- To service the infrastructure, they would need roughly ten times that figure.
Zitron’s framing: “The infrastructure being built and compute commitments being made are being done so at a level that demands that generative AI and AI compute generate over $2 trillion in annual revenue by 2030.”
That is the part that travels. The rest of the post is supporting evidence.
The data backing the case
Zitron walks through a series of recent disclosures that, taken together, make the slowdown argument hard to dismiss (estimating):
| Signal | Figure | Source |
|---|---|---|
| Planned new data-center capacity | 190 GW | Sightline Climate, Feb 2026 |
| Cost per GW (NVIDIA-reference build) | $80–100B | Jensen Huang, public remarks |
| Anthropic compute commitments | $330B | The Information |
| OpenAI 2026 compute spend | $50B | Bloomberg, May 2026 |
| OpenAI projected burn through 2030 | $852B | Zitron analysis |
| Anthropic projected 2029 revenue | $174B | The Information |
| OpenAI projected 2029 revenue | $184B | Author compilation |
| Combined market share | 89% of all AI startup revenue | The Information |
| Single enterprise token overspend | $500M / month | Zitron source reporting |
| Companies with zero AI cost visibility | 22% | KPMG, via Zitron |
Why it matters: every one of these numbers is sourced to a public disclosure or a major reporter. None of them are speculative. The unusual part is putting them in the same paragraph — which is what Zitron is doing, and what the keynote circuit is not.
What changed in Q1 2026
The pivot point in Zitron’s narrative is small but specific: Anthropic and OpenAI moved customers from flat-rate subscriptions to token-based billing in Q1 2026. Within two months, three things happened simultaneously:
- Enterprise CFOs started seeing line-item AI costs for the first time. KPMG’s most recent survey says 22% have no visibility into AI spend at all, and another 50% have only “some” visibility.
- Several large customers installed hard spending caps after sticker-shock incidents — one company, according to Zitron’s reporting, was burning $500 million a month on tokens before its CFO noticed.
- The financial press picked up the story. The phrase “AI ROI” started appearing in earnings calls without scare quotes.
Token billing is not the cause of AI’s economics problem. It is the moment the economics became visible.
What’s overstated
A fair read of the post requires pushing back on three things.
The $2 trillion target is a ceiling, not a hard requirement. Zitron’s math assumes the full 190 GW gets built on schedule and at NVIDIA-reference prices. Both assumptions are soft. Sightline Climate itself flags that 30–40% of announced capacity is likely to slip or shrink. Per-GW costs are dropping as hyperscalers move to liquid cooling and 800-volt DC distribution.
The “circular economy” framing is partial. Yes, NVIDIA’s revenue is heavily concentrated — the Q1 2027 earnings filing (Zitron previews it in the post) puts 54% of revenue at three clients. But NVIDIA, OpenAI, and Anthropic are not the entire AI economy. AWS, Azure, and Google Cloud are absorbing meaningful inference workloads that never touch the frontier labs. Mid-market enterprise AI — finance, legal, healthcare workflows — has shipped real revenue without making headlines.
Token overspend ≠ AI failing. A company burning $500M/month on tokens with no cost visibility is a procurement failure, not an AI failure. The same companies were burning that on cloud compute in 2014 before FinOps tooling caught up. Cost visibility tools for AI are now a Series-A category of their own.
Why the contrarian view is gaining ground
The reason the post resonated is not just the numbers. It is the timing. The token-billing rollout coincided with two other shifts that are harder to quantify but easy to feel:
- Generational jumps at the frontier are reportedly shrinking. Each new GPT, Claude, and Gemini release in 2026 has been positioned more around reliability, cost, and agentic plumbing than raw capability gains. That is a different marketing register from 2023–2024.
- Capex commentary on hyperscaler earnings calls has cooled. “Demand outpaces supply” — the phrase that defined four quarters of earnings — has started getting paired with ROI qualifiers.
None of that disproves the AGI roadmap. It does change the tone in which the roadmap is sold. Combined with the cost-visibility story, it is enough to give a skeptic post real traction.
What this means for builders and operators
If you ship a product that depends on frontier-LLM economics, the realistic 2026–2027 takeaways:
- Costs will keep falling per token, slowly. The 5–10x per-year compression of 2023–2025 is over. Plan for 2x annual cost compression at the frontier, faster on open models.
- Inference economics matter more than training economics. Margin sits in routing, caching, and small-model fallback. Build the abstraction now if you have not.
- AI procurement is becoming a board-level line item. That changes who you sell to and how. Expect annual contracts with negotiated rate caps to replace usage-billed credit cards by mid-2027.
- The hype-vs-data gap is your editorial advantage. Anyone shipping product-grade AI that demonstrates actual ROI will get an attention premium against the “AGI by 2027” pitch.
For more context, see our coverage of the 2026 AI capex wave and the GPT-5.5 launch reception.
What to watch next
- Whether OpenAI or Anthropic disclose gross margin on their API businesses in the next earnings cycle. They have avoided that line item for two years.
- Whether the 190 GW number gets revised in the next Sightline Climate update.
- Whether NVIDIA’s Q2 2027 filing confirms the 54% client-concentration figure — that is the single biggest market-fragility signal.
- The next viral skeptic post. Zitron is not alone — Gary Marcus and Ed Conway have been writing in adjacent territory. Watch for them to be cited in earnings calls.
The post that lit up Hacker News on Sunday is not the obituary. It is the first widely-shared piece that puts the infrastructure math in plain view. Whether the industry hits the $2 trillion line or not, the conversation has already shifted. The hype premium that funded the last three years of AI is going to be priced like every other premium — against the numbers it can actually produce.
FAQ
Who wrote “AI Is Slowing Down”?
Ed Zitron, publisher of the Where’s Your Ed At newsletter and one of the most-quoted AI skeptics of the cycle. The post landed on June 8, 2026 and trended on Hacker News within hours.
What is the $2 trillion number?
Zitron’s calculation: 190 GW of planned data-center capacity × $80–100B per GW = $9.5–15T in infrastructure commitments. Servicing that capex over a decade requires roughly $2 trillion in annual AI-compute revenue by 2030. Current trajectory points to around $358B combined for OpenAI + Anthropic by 2029.
Where do the underlying numbers come from?
Sightline Climate (Feb 2026) for capacity; Jensen Huang’s public remarks for per-GW cost; The Information for Anthropic compute commitments and revenue projections; Bloomberg (May 2026) for OpenAI’s 2026 compute spend; KPMG for enterprise visibility data.
Is the AI bubble bursting?
Probably not in the “tomorrow” sense. The honest read is that the funding model is under stress — token billing exposed enterprise cost-blindness, model improvements per generation are shrinking, and capex commitments outpace any plausible revenue path. None of that means the technology stops working. It means the business model gets repriced.
What changed with token-based billing in Q1 2026?
OpenAI and Anthropic moved enterprise customers off flat-rate subscriptions onto per-token billing. CFOs got real cost visibility for the first time. Several large customers installed spending caps after sticker-shock incidents — one was reportedly burning $500M/month before anyone noticed.
Are frontier model improvements actually slowing down?
The qualitative signal is that 2026 releases have been positioned around reliability, cost, and agentic features rather than raw capability gains — which is a different register from 2023–2024. Whether the underlying benchmark deltas have actually shrunk depends on which benchmark you read; published numbers vary.
What should AI builders do about this?
Three things: design for 2x annual cost compression (not 5–10x), invest in routing and small-model fallback now, and prepare for procurement-driven sales cycles by mid-2027. Anyone with provable ROI gets a market premium against the AGI-pitch competition.
Why is the contrarian view gaining ground now?
Timing. Token billing exposed the cost picture for the first time in Q1 2026, marketing for new model releases shifted away from raw capability claims, and capex commentary on hyperscaler earnings calls started pairing with ROI language. Each of those is small on its own. Together, they give the skeptic case real traction.
Want the AI economics dispatch as it lands each week? Subscribe to the TechDaily360 weekly briefing.