Apollo: AI Still Hasn't Paid Off for the Other 493 S&P 500 Companies — Tokens Become a Cost Again, Not an Engine
🔎 The Note That Silences the Hype
Torsten Slok, chief economist at Apollo Global Management, has just published a note that should be required reading for every executive board: "No Signs of AI Payoff Outside the Tech Sector." Dated October 24, 2026, it doesn't predict the future — it documents the present. Five charts, inflation-adjusted series since Q4 2019, and a scathing verdict: outside the "Magnificent 7," AI has produced no measurable financial return. Revenue per employee and margins for the S&P 493 are flat. Total factor productivity, as measured by the San Francisco Fed, isn't moving. And tokens, once hailed as the new oil, are becoming a cost center again that weighs on margins. The AI story, Slok concludes, "will play out in the S&P 493, not the Mag 7."
Key Takeaways
- Mag 7 vs S&P 493: revenue per employee ≈ $780k vs ≈ $527k (Q4 2019 prices), margins ≈ 30% vs ≈ 9-10% — the gap is widening, but only for 7 companies.
- Productivity: the San Francisco Fed's adjusted TFP remains at zero despite billions injected into GPUs and models.
- Extreme concentration: the top 10% of firms capture 99.5% of inference, 99% of neoclouds, and 95.6% of OpenAI + Anthropic payments (Ramp AI Index data).
- Circularity: 80.4% of the top 1% of AI spenders are backed by venture capital — money circulates in a loop between VCs, hyperscalers, and model providers.
- Token prices: Fable 5 at ≈ $10/M tokens (August 2026), twice the price of GPT-5.6 Sol — the cost curve isn't coming down for everyone.
- Key quote: AI "understands neither responsibility, nor trust, nor culture" — Slok, Apollo Daily Spark, Oct. 24, 2026.
- Counterpoint: Brad DeLong accepts the narrow comparison but argues that user benefits and new activities matter more than the profits of existing firms.
Recommended Tools
| Tool | Main use | Price (October 2026) | Ideal for |
|---|---|---|---|
| Ramp AI Index | AI spend tracking per employee/month | Free (public data) | CFOs, ops, AI strategy |
| Apollo Daily Spark | Macro/AI notes from Torsten Slok | Apollo subscription | Investors, strategists |
| Fable 5 (via API) | High-end LLM inference | ≈ $10/M tokens (August 2026) | Teams testing premium models |
| GPT-5.6 Sol (OpenAI) | Standard enterprise inference | ≈ $5/M tokens (August 2026) | High-volume production |
The 5 Charts That Hurt — Mag 7 vs S&P 493
Apollo's note rests on five inflation-adjusted time series, indexed to Q4 2019. The first chart shows revenue per employee: the Magnificent 7 (Apple, Microsoft, Nvidia, Amazon, Meta, Alphabet, Tesla) go from ~$500k to ~$780k. The S&P 493 stays stuck at ~$527k. Same trajectory for operating margins: ~30% for the Mag 7, ~9-10% for the rest. These figures are not projections — they are audited income statements, adjusted for inflation.
What this means: AI may have boosted productivity inside the seven giants that build and sell it. For the other 493, it's still an expense line, not a revenue line. Slok's note confirms it unambiguously: "zero financial payoff" for corporate America outside tech.
Total Factor Productivity: The Big Zero from the San Francisco Fed
The most stinging chart isn't an accounting one — it's macroeconomic. The Federal Reserve Bank of San Francisco publishes a utilization-adjusted total factor productivity (TFP) series. Since Q4 2019, the curve has been flat. Zero. Nada. No bump, no acceleration, not even a statistically significant tremor.
Yet capital expenditures (CapEx) for AI have exploded. Nvidia alone posted more than $100 billion in data center revenue in 2025. The hyperscalers (Microsoft, Amazon, Google) have deployed hundreds of billions in GPUs. If AI were a general-purpose technology (GPT in the economic sense), TFP should have moved. It hasn't.
Interpretation: either AI hasn't yet been deployed at productive scale outside tech, or its gains are being captured by the suppliers (models, cloud, chips) rather than by end users. Ramp's data lean toward the second explanation.
Tokens — from growth engine to cost center (Ramp data)
The September 2026 Ramp AI Index publishes figures that send chills down the spine: median AI spend per employee per month = $15.70. Top 10% = $813.43. Top 1% = $9,567.53. The distribution is more unequal than income in the United States.
Worse: the note "Circularity in AI Shows Up in Token Spending Too" from Apollo reveals that 80.4% of the top 1% of AI spenders are VC-backed. Money flows out of the LPs' pockets (pension funds, foundations), into the VCs, funding startups that buy inference from the hyperscalers and model providers, who in turn buy GPUs from Nvidia — whose shares are held by... the LPs. A closed loop.
The Ramp Economics Lab notes in August 2026 that "companies are hitting their AI spending limit." Fable 5, a premium model, charges ≈ $10 per million tokens — twice the price of GPT-5.6 Sol. The promise that "costs fall exponentially" only holds for commoditized models. At the frontier, the bill keeps climbing.
This is where tokens go back to being a cost center. Not a productive investment. A burden weighing on S&P 493 margins. Understanding LLM billing becomes a survival skill for CFOs.
The circularity of AI spending — 80% of the big payers are VC-backed
The circularity is not a conspiracy theory — it's accounting. Apollo's chart shows it: the top 1% of AI spenders concentrate a massive share of token volume. And 80.4% of that 1% are venture capital-backed companies.
Why it matters: these companies are not yet profitable. They burn cash to grow, and a growing share of that cash goes to tokens. When the VC cycle tightens (rising rates, difficult fund exits), token demand collapses. This is exactly what Slok calls a "painful repricing" — a warning sounded as early as October 22, two days before the note.
This circularity also explains why hyperscalers and model providers (OpenAI, Anthropic, xAI, Google) keep announcing record revenues: their best customers are funded by the same ecosystem that values their shares. 46% of AI tokens from American companies already go to Chinese models — a headlong rush that accelerates pressure on prices.
Brad DeLong's rebuttal — user benefits vs profits of existing firms
Brad DeLong, economist at UC Berkeley, responded on Digg: he "accepts the narrow comparison" from Slok (profits of existing firms, Q4 2019 adjusted series) but rejects the conclusion. His argument: the benefits to users and new activities matter more than the profits of existing firms.
Example: if AI lets a doctor save 30% of their time on paperwork, that gain doesn't show up in their hospital's revenue per employee (the hospital being an existing firm). It shows up in quality of care, patient access, and the creation of new services. DeLong cites the history of electricity: 20 years elapsed between the installation of motors and the surge in factory productivity.
My take: DeLong is right on principle, wrong on timing. Public-market investors (S&P 493) don't have 20 years. They have quarterly reports. And Apollo's note shows that the market's patience has a limit — the "painful repricing" isn't a threat, it's a deadline.
The Hacker News signal and the "painful repricing" warning
The Apollo note hit 300+ points on Hacker News within hours according to explainx.ai. That's not a trivial detail: HN is the break room for the engineers who build and buy AI. The fact that the note blew up there means that practitioners feel the disconnect between the marketing narrative and the reality of their cloud bill.
Slok had warned on October 22: "painful repricing coming." Two days later, he released the charts. The timing is no coincidence. Markets have priced in the idea that AI = infinite growth. The data says: AI = concentrated costs, concentrated gains, VC circularity.
This repricing will first hit the "neoclouds" (CoreWeave, Lambda, etc.), whose 99% of business comes from the top 10% of firms according to Ramp. Then the model providers, whose 95.6% of OpenAI+Anthropic payments come from that same top 10%. Finally, the S&P 493 companies that overinvested without a clear use case.
What This Changes for Your AI Strategy
Stop buying tokens by the pound. The median of $15.70/month/employee (Ramp, Sept. 2026) masks a reality: most companies don't need GPT-5.5 or Claude Opus 4.7 for 90% of their use cases. Open models already power 56% of Vercel's tokens — the quiet migration away from OpenAI/Anthropic is underway, driven by cost and control.
Audit your use cases by ROI, not by FOMO. If a workflow doesn't generate additional revenue or measurable cost savings within 6 months, cut it. Tokens have become a cost center again. Treat them as such: negotiate, pool resources, and switch to self-hosted when volume justifies it.
Watch the S&P 493, not the Mag 7. That's where real adoption will happen. The 493 companies that have not yet seen their productivity move are the real addressable market. Those that find the path to "payoff" will become the next industry leaders. The rest will fund Nvidia's, Microsoft's, and Google's R&D.
❌ Common mistakes
Mistake 1: Confusing adoption with value creation
What's wrong: "We rolled out Copilot to 5,000 seats, so we're doing AI."
The solution: Measure the productivity delta per workflow. If revenue per employee hasn't moved after 12 months, you're paying for a subscription, not a transformation.
Mistake 2: Ignoring spending concentration
What's wrong: Benchmarking your AI budget against the Ramp median ($15.70/month/employee).
The solution: The median is misleading. The top 1% spends 600× more. If you're not in the top 10%, your budget should aim for efficiency, not raw power.
Mistake 3: Believing costs go down for everyone
What's wrong: Projecting the GPT-4 → GPT-4o → GPT-5 cost curve onto your premium models.
The solution: Fable at $10/M tokens (August 2026) proves that the cutting edge is getting more expensive. Negotiate volume commitments or migrate to open models (Llama 4, Nemotron, Qwen) for non-critical workloads.
❓ Frequently Asked Questions
Is AI really making money for anyone outside Big Tech?
Not for the shareholders of existing firms, according to Apollo's Q4 2019 adjusted series. But DeLong is right: gains do exist for end users (patients, customers, employees) and for new companies born with AI. They just don't show up in the P&L of the S&P 493.
Should you stop investing in AI?
No. But invest like a rational buyer, not like a VC. Every dollar spent on tokens must have a traced path to revenue or cost savings. Manage by quarterly ROI, not by "AI strategy."
Will the "painful repricing" cause a crash?
Not a 2000-style crash. A brutal reallocation: capital toward proven use cases, sharp cuts to experimentation without metrics. Neoclouds and model providers with VC-heavy client bases will be hit first.
How do you break out of the VC-token circularity?
Diversify your providers. Test self-hosting for stable workloads. Demand cost SLAs per token, not just latency. Make LLM billing an internal competency, not an opaque line item at your cloud provider.
✅ Conclusion
Apollo's note doesn't kill AI — it holds it accountable. $780k of revenue per employee for the Mag 7, $527k for the S&P 493. 30% margin vs 10%. Zero macro productivity. 80% of big spenders VC-backed. Tokens have become a cost again. The AI story will play out in the S&P 493. The 493 companies that rise to the challenge will become the new giants. The rest will have funded the lesson. It's up to you to pick your side — and your token bill.