AI capex is not a single-year P&L problem

Why 2026 probably brings digestion and remix, not collapse, and what that means for data center buildouts

When you read headlines about the AI capex boom, it can feel dizzying. The numbers are staggering: hundreds of billions in spending commitments, entire supply chains re-wired for GPUs and high-bandwidth memory, data center power demand outstripping cities. Then you see a tweet like Josh Wolfe’s, which compares that spending to the relatively modest revenues attributed to AI, and you might wonder: is this all just a bubble?

Tweet is here. Thread is here.

The answer isn’t simple. There is a mismatch between how much money is going out the door and how much revenue explicitly tied to AI is coming back in. But that comparison misses the structure of capex, how hyperscalers account for value, and the timing of returns. The story isn’t one of collapse; it’s more a question of digestion and remix. Think of 2026 as a year where the big players slow the growth rate, sweat the assets they’ve already built, and redirect dollars toward training-class clusters, memory-rich pods, and custom silicon.

To orient ourselves, let’s start with the basics of how capex works, why the revenue denominator is misleading, and what shifts in the mix are coming next.

Capex vs “AI revenue”: different currencies

Capital expenditures are long-lived assets. When Microsoft, Meta, or Amazon spends billions on GPUs and data centers, those servers, racks, and power contracts are designed to generate cash flows over 4-6 years (sometimes longer). “AI revenue,” as it’s often cited, is a one-year flow number, and usually a very narrow one.

• Indirect benefits don’t show up cleanly. Meta’s Llama models improve ad targeting, boosting click-through rates and yields, but that shows up as “Ads” not “AI.” Microsoft’s Copilot raises seat ARPU across Office, but that rolls up into Productivity. Amazon uses AI to cut logistics costs and churn in AWS, but those savings aren’t labeled “AI revenue.”

• Capex buckets are broad. Hyperscaler capex covers more than GPUs: logistics, infrastructure, general cloud, networking, even moonshots like VR. Treating every dollar as if it were spent solely on AI skews the comparison.

This is why comparing $560B in capex to ~$35B of “AI revenue” yields a scary ~6% figure that isn’t really ROI. The real question is: given the useful life of the assets¹ and the cost of capital, what level of annual cash generation do these assets need to justify themselves?

A better yardstick: annual cash flow vs useful life

Let’s reframe. Assume ~$500-600B in AI-relevant PP&E by end of 2025. Accelerators depreciate in 4-6 years; the corporate cost of capital is ~8-10%. That implies required annual cash generation in the range of $120-180B to clear the hurdle. Even if only 60% of capex is truly AI-specific, that still means ~$70-100B per year in incremental cash flow needs to come from AI-enabled activity.

A simple model explains the math²: PV = CF * (1 - (1+r)^-n) / r

PV = $500-600B

r = discount rate (0.08 to 0.10)

n = useful life (4-6 years)

If we plug in the numbers:

• 4 years, 10%: Factor = 0.3156 —> Required CF ~= $550B * 0.316 ~= $174B/year

• 6 years, 8%: Factor = 0.216 —> Required CF ~= $550B * 0.216 ~= $119B/year

Across the $500-600B range, you get ~$120-180B/year.

That’s the bar, not $35B of “AI revenue”. And that bar isn’t impossible. If ad platforms add a few hundred basis points of efficiency, if enterprise software bundles AI into high-margin seats, and if cloud providers sell managed services on top of their GPU fleets, the incremental cash flow adds up.

On-device inference trims the cloud

One structural shift is undeniable: a growing share of inference will run on devices, not in the cloud. Why?

1. Latency and privacy. Many “good enough” tasks, like autocomplete, summarization, lightweight image generation, run faster and safer on device NPUs.

2. Unit economics. If Apple can keep inference local to the iPhone, it avoids cloud serving costs entirely. Multiply that across billions of devices, and the incentive is obvious.

3. Silicon variety. Custom TPUs and NPUs optimized for smaller, distilled models are eating the middle of the inference market.

What does that leave the cloud? Heavy training and fine-tuning, high-fidelity multimodal generation, collaboration or retrieval-rich workflows, and enterprise-grade compliance. That’s still large, but it means hyperscalers will tilt spend toward training-class clusters and bandwidth-dense, memory-heavy nodes, rather than ever-larger inference farms.

2026 as downshift and remix

So what does 2026 look like?

• Downshift. The explosive growth of AI-branded capex slows. CFOs will gate new clusters on utilization thresholds: is the fleet 70% full within 90 days? 85% by 180? If not, no new dollars. This creates volatility for suppliers like Nvidia: feast in 2024-25, then a digestion period.

• Remix. Dollars tilt from general inference toward (a) training pods, (b) custom ASICs, (c) disaggregated memory/interconnect. More emphasis on modular design so fleets can swing between training and inference. Asset lives stretch: older GPUs are pushed down to inference, vision, or VDI workloads.

Data centers still get built, but the shape changes:

• Fewer greenfield halls built purely for open-ended inference. More densification on existing campuses, designed for power density and fabric efficiency.

• Power remains the limiting reagent. Expect more long-dated PPAs, siting near firm power supply and behind-the-meter generation.

• Square footage grows slower.

• MW capacity planning dominates.

Company snapshots

• Microsoft: AI revenue is bundled. The OpenAI arrangement looks low‑margin, but the prize is Azure consumption and per‑seat Copilot uplift.

• Amazon: Bedrock and managed AI services grow slowly, but AI permeates Ads and Retail ops. Reserved capacity contracts better align cash flows with asset life.

• Meta: Even without a direct AI SKU, a few basis points of ad efficiency across its base repays billions of spend. Open‑sourcing Llama also pushes inference to the edge, lowering Meta’s own serving bills.

• Google: TPU investment cushions Nvidia dependence. The challenge is cannibalization: AI‑enhanced search has to repay itself in ad yield.

• Tesla: GPU spend is a levered bet on FSD. If autonomy hits, returns look software‑like; if not, the capex was oversized.

Practical playbook for 2026

1. Capex gating by workload. Fund only against clear utilization models and per‑GPU gross margin projections.

2. Design for repurposing. Modular, memory‑rich pods that can swing between workloads.

3. Monetize via bundles and SLAs. Raise seat ARPU, offer tiered latency/privacy packages, keep metered APIs for developers.

4. Treat power as part of ROIC. GPUs without MWs are stranded assets; lock supply early.

5. Hedge obsolescence. Mix of own vs. lease, vendor buybacks, and active secondary markets.

The bottom line

Comparing capex and “AI revenue” one‑for‑one is misleading. The right lens is multi‑year cash generation versus asset life and cost of capital. By that standard, 2026 is likely to be a year of discipline: slower slope, stricter gating, and a remix toward training fleets, memory‑centric designs, and custom silicon. Cloud AI doesn’t vanish, but it specializes. On‑device takes a bigger bite out of inference.

For data center builders, that means fewer generic halls and more dense, power‑first expansions. For vendors, it means volatile order books and pressure to help customers extend asset life. For investors, it means stop counting just the AI line and start tracking the indirect engines, including ads yield, per‑seat uplifts, and PaaS margins, that really determine whether this capex cycle pays off.

---

If you enjoy this newsletter, consider sharing it with a colleague.

I’m always happy to receive comments, questions, and pushback. If you want to connect with me directly, you can:

• follow me on Twitter,

• connect with me on LinkedIn, or

• send an email to dave [at] davefriedman dot co. (Not .com!)

1. Incidentally, the question of the useful life of GPUs is an interesting one. See Jim Chanos’ tweet here. FASB provides a standard for useful life, which is used for depreciation purposes, but thermal cycling and new chip generations likely shorten a GPU’s functionally useful life.

2. I can’t figure out how to render equations properly in Substack, so this plaintext rendering will have to suffice.