AI Infrastructure Economics v2

The Little House in the Cloud, Measured

The pitch (Emad Mostaque on Raoul Pal's Journey Man, “The AI Job Shock Is Coming,” recorded September 28, published October 8; transcript ~45:31–47:00): cloud agents each need a VM, a VM is CPU cores plus RAM, OpenAI, Anthropic and Meta will all host agents around the clock, the suppliers are “AMD and Intel,” so “CPUs will be the new shortage.” He also says the DRAM gap “will resolve in six months.” It's a forecast on a promotional podcast, so I tested it rather than repeating it. Half of that is right and half is a units error.

Right: the server CPU market is genuinely short in 2026. Intel's own 10-Q says server demand exceeded supply because of internal constraints and expects substrate and memory shortages to run into 2027 (Intel 10-Q). Intel CEO Lip-Bu Tan says Intel can serve only about half of what customers are asking for (Motley Fool). TrendForce puts server CPU lead times at 25–30 weeks against a normal 16–20 (XenoSpectrum), and a July TrendForce note said CPU shortages were delaying server assembly so badly that cloud providers were sitting on DRAM they couldn't install (TrendForce).

The units error: it is not one CPU per agent, or anything close. Nvidia sizes its 256-CPU Vera rack for more than 22,500 concurrent environments, roughly one sandbox per physical core (Nvidia). Firecracker microVMs add about 5 MiB of overhead each, and AWS has demoed 4,000 of them on one 36-core server (Firecracker spec, demo). Lambda runs production oversubscription as high as 10× (NSDI '20). So 100 million concurrent agents at one per core is 100M cores, about 1.25M 80-core sockets. That's roughly 4% of a year's server CPU supply. Real, but not a shortage by itself.

Memory: Emad Says Six Months. Micron Says 2028.

Emad's other forecast is that the DRAM gap “will resolve in six months.” Nobody who makes DRAM agrees. TrendForce had server DRAM contracts up 13–18% QoQ in Q3 and conventional DRAM up another 10–15% in Q4, with server DRAM still undersupplied (July, Sept 30). New fabs don't contribute meaningfully until late 2027–2028. Micron's September prepared remarks say conditions could tighten further in 2027–2028 (Micron, Computerworld). The model defaults to “tight through 2027.” Treat six months as the optimistic tail, not the base case.

The agent link is real, though. TrendForce says general-purpose x86 servers with RDIMM are the primary memory platform for agentic workloads (TrendForce). Micron's own blog makes the same argument: every agent environment carries its own memory footprint (Micron). At Muse's ~8 GB per user, 100M users fully provisioned is 800 PB of DRAM. CPU and DRAM shortages also feed each other: in July, CPU shortages left cloud providers holding DRAM they couldn't install. When CPU supply improves, that DRAM gets absorbed and the memory pull strengthens.

The Bear Layer: Prism ML, DeepSeek, and Bytes per Token

• Prism ML Ternary Bonsai 27B (July 14): Qwen3.6-27B in ~1.7 effective bits per weight. 5.9 GB ideal (7.2 GB deployed) vs 54 GB at FP16, at 94.6% of FP16 benchmark score. That's the “~5% loss” claim, and it checks out on Prism's own evals (Prism ML, model card).
• Ternary Bonsai 2 27B (September): Qwen3.8-27B at 5.9 GB, 83.9 vs 85.4 for FP16, or 98.2% retention, vendor-reported (Prism ML).
• DeepSeek V4.1-Flash (~Sept 10): company-reported ~75% less HBM and ~87.5% less SSD for KV cache, not weights (Insider Monkey, Bloomberg). Google's TurboQuant KV compression (March) is the same idea (Google Research).

Power: Still the Slowest Thing to Build

v1 called power the binding constraint. Eight months later that holds, with a twist: the money showed up but the electrons are late. Goldman's scheduled US additions were 13.6 GW for 2026 and 36.3 GW for 2027, but it estimates only 50–60% arrives on time. Its October update raised end-2026 capacity to 64 GW and trimmed end-2027 to 90 GW. Interconnection still runs 3–4 years in major markets and up to 7 in Northern Virginia, and only ~5 GW of the 16 GW announced for 2026 delivery was under construction (TD Economics, Allianz). Behind-the-meter gas is the escape hatch: Cleanview sees ~10 GW of it possible in 2027, with a ~5 GW downside case (Cleanview).

Grading the February Model

v1 published February 27, 2026. Here's every name it rated, from the February 27 close to the October 12 intraday price (~11:10 ET). No cherry-picking, no “directionally correct.”

What v1 Got Right, Wrong, and Missed

• Right: demand outran supply. Nvidia's data center revenue hit $89.0B last quarter (+117% YoY), with total revenue guided to $108B (Nvidia). DELL and MU as the under-owned beneficiaries.
• Right on fundamentals, wrong on stocks: power. The constraint is still real, but power names went sideways once yields spiked and the market started pricing financing risk.
• Wrong: treating the GPU as the only scaling unit. v1 rated the two companies selling the other half of the AI data center, CPUs and general-purpose servers, as Holds.
• Missed: Intel. A CPU shortage is first a pricing event, and the company with the most supply-constrained capacity gets the most pricing power.

Hidden Gems: The DELL-Analogs of the CPU Layer

v1's best call was the boring box-maker everyone underweighted. The question for v2: where does a CPU-and-memory squeeze show up in earnings before it shows up in the multiple? The obvious names (AMD, Intel, Micron) have already moved 2–3×. These five haven't, or have round-tripped. Each has a specific mispricing, not just “small and cheap.”

Prices: Yahoo Finance, October 12, 2026 ~11:10 ET. Consensus targets: Yahoo Finance analyst mean. Forward P/E: Yahoo.

The Thesis in One Paragraph

The AI buildout is now short on everything at once, and the CPU layer is the one the market mispriced. In February, v1 said GPUs, tokens and power couldn't keep up through 2027–28. That held: Nvidia's data center revenue was $89.0B last quarter, +117%, with total revenue guided to $108B. What v1 missed is that agents turned the CPU from a commodity host into a scarce input. Intel says it can meet about half of what customers ask for, AMD's 2027 Venice volume is reportedly allocated, lead times run 25–30 weeks, and Intel's server ASPs are up 48%. But “CPUs are the next shortage because every agent needs a VM” gets the mechanism wrong. Agents are packed roughly one per core, so even 100M concurrent sandboxes are ~4% of a year's sockets. The shortage is driven by the CPU:accelerator attach ratio (1:4–1:8 heading toward 1:1–1:2), a general-purpose refresh the agents triggered, and supply bottlenecks in Intel capacity, substrates and TSMC N2. My base case: tight through 2028 (92–93% coverage), a brief easing as Intel 3, 18A, Venice and Vera ramp, then tighter again in 2029–30 as the attach ratio keeps climbing.

What Changed Since February

The Three Sources of CPU Demand

1. The general-purpose refresh (biggest, least glamorous)

Dell's traditional servers and networking did $10.53B last quarter, +122%. Jeff Clarke said agentic workflows need additional CPU capacity, but most of the growth is existing customers modernizing (Dell). HPE's server revenue rose 35%. Lenovo's ISG rose 98%. The catch, from Dell's own 10-Q: traditional-server growth was primarily higher ASPs, driven by memory-cost inflation. Intel shows the same pattern: server ASP +48%, units +9%. The refresh is real, but in 2026 it shows up mostly as price.

2. Accelerator-attached CPUs (the swing factor)

GB200 NVL72 pairs 36 Grace CPUs with 72 GPUs (1:2). Fleet-wide, TrendForce puts today's AI data centers at 1:4–1:8 and sees agentic deployments heading to 1:1–1:2. With ~18M accelerators shipping in 2026 and ~25M forecast for 2027 (DigiTimes: 25.95M, with ASICs overtaking GPUs; Epoch tracks the installed base), each 0.1 of ratio is ~2M sockets. That's why the model's ratio sliders move the CPU verdict far more than any agent-count slider.

3. Agent sandboxes (the headline, the smallest)

Muse's per-user VM, Claude Code's 4-vCPU cloud sessions and Codex's 2–4 vCPU tasks are real and growing. But concurrency is a fraction of users, and packing is ~1 sandbox per core at Nvidia's own design point. The model's default of 40M daily agent users at 3 sandbox-hours each yields ~15M peak concurrent sandboxes in 2026, only ~0.2M sockets. Sandboxes reach socket-level significance (5%+) only with hundreds of millions of concurrent agents or dedicated multi-core VMs. Where they do bite is RAM (8–16 GB each) and who owns the capacity: Meta bought tens of millions of Graviton5 cores, and Anthropic is renting CPUs from Akamai.

Supply: Why It's Short, and Why It Eases

• Intel: demand exceeds supply on internal constraints. Substrates and memory stay short into 2027. Intel 3 is the key Xeon node (a €5B Ireland expansion), and 18A output ran 25% above target in Q2 (10-Q, call).
• AMD: Venice is the first HPC product on TSMC N2, ramping since May (AMD). Lisa Su: supply was tight on unforecasted demand, and 2027 should be better than 2026.
• New entrants: Nvidia Vera on N3 (shipping since August), Arm's AGI CPU (broad availability 2H26), and Graviton5 / Cobalt 200 / Axion inside the hyperscalers.
• The tell that it's easing: TrendForce's September 30 DRAM note describes improving server CPU availability, which is letting cloud providers finally install the memory they bought (TrendForce). I found no evidence that GPU wafers are crowding Venice off N2.

Counter-Theses (Addressed, Not Dismissed)

“It's a price event, not a volume event.”

Mostly true for 2026. Intel units +9% against ASP +48%, and Dell's growth is ASP-led. That's why I don't chase Intel at 48× forward: scarcity pricing is the most perishable kind of earnings. It doesn't break the thesis, because the attach-ratio shift is a unit story that compounds into 2028–30. It does change which stocks to own: suppliers whose content per server rises (memory interface, CXL, server OEMs) beat suppliers whose price per chip rises.

“Double ordering. This ends in a 2027 inventory hangover.”

Plausible, no evidence yet. Classic shortage behavior: 25–30 week lead times breed duplicate orders, and the CPU-without-DRAM mismatch in July looked like it. But reporting says Intel entered 2026 with depleted buffer inventory, and demand exceeded supply in both reported quarters. Watch for lead times compressing below 16 weeks and Intel price cuts. That is the signal to trim the CPU-scarcity names.

“Agents run in shared containers, not VMs, so CPU demand is overstated.”

Correct, and already in the model. This is the argument against the sandbox lens, which I already score as minor. It doesn't touch the attach-ratio or refresh lenses. Slide packing to 8 per core and the CPU verdict barely moves.

“Efficiency kills memory demand” (Prism ML, DeepSeek).

A real long-run risk, oversold short-run. A 27B model in 6 GB at 95–98% quality is a genuine threat to cloud inference of small models, and it's the strongest argument that some agents move on-device. But it shrinks bytes per weight and bytes per token, not bytes per agent environment, and so far cheaper tokens have bought more tokens. Memory suppliers say tight through 2027–28. The model's haircut slider is where you express disagreement.

“It's not AMD and Intel, it's Arm and Nvidia.”

Half right, and it's the most important nuance in this report. Arm-based chips are ~half of the compute shipped to top hyperscalers and 13.6% of server CPU units. Graviton has been the majority of AWS's new CPU capacity for three years. Nvidia is selling CPU-only Vera racks. The x86 duopoly is selling into a shrinking share of the most important buyers. That's why AMZN and NVDA rank above INTC here, and why Arm's royalty stream is worth watching even though the stock isn't cheap.

“Rates and financing stop the buildout.”

This is the one that has actually hurt stocks. Oracle spent $28.5B of capex against $23.1B of operating cash flow last quarter (Oracle). OpenAI's run-rate was reported ~$20B below circulating estimates (CNBC). CoreWeave is issuing converts and ATM equity. Higher yields make data center debt harder to raise (CNBC). The buyers that matter for CPUs (Amazon, Meta, Microsoft, Google, Anthropic) are funding from cash flow or long contracts. Leveraged intermediaries are the weak link, which is why CRWV and ORCL are on the avoid list.

What Could Break This Thesis

The Case for Each Rating

NVDA: Strong Buy, PT $300–320

The cheapest it has been in this cycle: ~14× forward for $96.2B of quarterly revenue (+106%) and a $108B guide. Vera lets Nvidia sell into the CPU layer it previously gave away, with OCI deploying hundreds of thousands of them. The market is pricing peak margins and China. I'm pricing a company that now sells both halves of the AI rack.

0992.HK Lenovo: Strong Buy, PT HK$42–46

See Hidden Gems. #1 x86 server shipper per its IDC citation, ISG +98%, at 10.5×. The single best expression of “someone has to build the boxes” that hasn't re-rated.

AMD: Buy, PT $680–720 (upgraded from v1's Hold)

v1's biggest miss, owned. Server CPU revenue guided +80% in 2H26 and +70% in 2027, Venice on N2 with 2027 volume reportedly allocated, and 34.5% x86 unit share and rising. Not a Strong Buy at 37× with up to 320M penny warrants promised to OpenAI and Meta (Motley Fool), and Vera and Arm chewing the hyperscaler socket. This is a share-gain story, which outlasts scarcity pricing.

AMZN: Buy, PT $310–330

The stealth CPU winner. Graviton has been most of AWS's new CPU capacity for three years, Meta signed for tens of millions of Graviton5 cores, and Anthropic's $100B+ AWS commitment includes Graviton. You get the agent-CPU landlord at 25×.

TSM, DELL, MU, VRT: Buy

TSM fabs both Venice (N2) and Vera/Rubin (N3), with no sign of CPUs being crowded out, and has wafer pricing power. DELL: traditional servers +122% and a $95B AI backlog at under 20×, though ASP-led growth means memory-cost reversals cut both ways. MU: Micron says tight into 2027–28. The 4.8× forward multiple is the market saying “peak,” so size for cyclicality. VRT: power is still the slowest layer to build, and the stock is 37% off its high.

INTC: Hold (not rated in v1)

The purest scarcity-pricing play, and that's the problem. Units +9%, ASP +48%, 48× forward, foundry still losing ~$2B a quarter. TrendForce already sees CPU availability improving. I'd rather own the content-per-server names than the price-per-chip name at this point in the cycle.

ARM: Hold

Data center royalties more than doubled and Arm is ~half of hyperscaler compute. But selling its own silicon dilutes a royalty business with ~97% gross margins (Yahoo), and 84× forward still prices perfection after a 43% drawdown.

AVOID: CRWV, ORCL, OKLO

The financing layer is where this cycle is breaking first. Oracle's quarterly capex exceeded operating cash flow by ~$5.4B, and it is negotiating off-balance-sheet chip financing. CoreWeave is issuing converts and ATM equity. Oklo is pre-revenue, sitting on its 52-week low. Same demand, worst balance sheets.

Portfolio Construction

How This Model Works

Four supply layers, each modeled independently for 2026–2030, then compared. Coverage = supply ÷ demand. The verdict takes the lowest coverage of any layer in 2026–27 and bands it: ≥100% supply keeps up, 90–100% tight, 75–90% shortage, <75% severe.

Layer 1: Tokens (the v1 spine, re-anchored)

• Demand = monthly tokens × growth^t ÷ seconds per month × peak-to-average. Default 8 quadrillion/month in 2026, against Google alone at 3.2 quadrillion/month across its surfaces (May) and Google APIs at ~22B tokens/minute (Q2).
• Supply = installed H100-equivalents × inference share × tokens/sec per H100e × (1 + efficiency gain)^t × utilization, then capped by power coverage, since chips without power serve nothing. The default fleet of 33M H100e scales Epoch AI's tracked 14.4M (43% coverage). Analyst estimates run from ~31.6M (Epoch, Sept) to ~50M (year-end).

Layer 2: CPU sockets (new in v2)

• General-purpose: 28M sockets in 2026, growing 6%/yr. Derived from ~20M server units (DigiTimes, via Lattice) at ~1.5 sockets per server, minus the AI-attached portion. Planning estimate, not a published figure.
• Accelerator-attached = accelerators shipped × CPU:accelerator ratio, interpolated linearly from the 2026 to the 2030 setting. Defaults: 18M accelerators (forecasts range 16.9–19.4M) growing 40%/yr (DigiTimes 2027: 25.95M). Ratio 0.25 (1:4) to 0.6 (~1:1.7), inside TrendForce's 1:4–1:8 → 1:1–1:2 path.
• Agent sandboxes = daily agent users × (sandbox hours ÷ 24) × parallel agents × peak ratio ÷ sandboxes per core ÷ cores per socket. Cores per socket rise 12%/yr (Venice to 256 cores, Graviton5 192, Vera 88).
• Supply: 30M sockets in 2026 growing 20%/yr. Lisa Su says 2027 supply is better than 2026; Morgan Stanley’s reported Venice estimate is ~6.75M units in 2027 vs ~1.25M in 2026 (TechTimes); Intel 3 and 18A are ramping.

Layer 3: DRAM bits

Indexed to 2026 demand = 1. Supply starts at 98% of demand (TrendForce: undersupplied in 2026) and grows 20%/yr. Demand grows 25%/yr less a memory-efficiency haircut. Sandbox RAM is shown separately because it's small against total bits (server DRAM is tens of exabytes a year), but it's the part that scales directly with agents.

Layer 4: Power

Need = installed H100e × all-in kW per H100e ÷ (1 + perf-per-watt gain)^t. Available: 45 GW of global AI data center power in 2026, growing 35%/yr. US data center demand alone runs 41 GW (2026) → 66 GW (2027) per Goldman, and only part of that is AI.

What the model does NOT do

• No price elasticity. Shortages raise prices, which suppresses demand. The model shows the gap that pricing has to close; it doesn't predict the price.
• No substitution between layers, e.g. agents moving on-device under Prism-style quantization. Use the agent-users and efficiency sliders to approximate it.
• No geography. The power layer is global; the binding constraint is often a single substation.
• Sockets are not cores. A 256-core Venice and a 64-core Xeon count the same. Rising cores per socket is the main reason sandbox demand stays small in socket terms.

Key sensitivities (which slider moves the verdict most)

1. CPU:accelerator ratio, 2030: the CPU layer's dominant input. At 1.0 the CPU layer is in shortage by 2029.
2. Token demand growth vs software efficiency gain: the token layer's whole story.
3. Power growth vs perf-per-watt: set power growth to 20% and power binds everything by 2028.
4. Packing: matters only below 0.5 sandboxes per core (dedicated multi-core VMs) and with 200M+ daily agent users.

Research process

The first pass was 64 Perplexity queries from a background job that timed out mid-research. This version re-ran 30 multi-engine searches through wide-search (Grok with web + X, ChatGPT search, and Perplexity in parallel), preferring claims two or more engines agreed on and primary sources. Where engines disagreed (the DeepSeek selloff attribution, the origin of the 4–40 ratio), the report says so. Every price and market cap is a Yahoo Finance quote pulled October 12, 2026 at ~11:10 ET. Every linked source came back from a search engine result; none were constructed. Things I could not verify were dropped or labeled: Emad's six-month DRAM call, Meta's exact Muse VM spec (secondary reporting), Lenovo's #1 x86 shipment rank (Lenovo citing IDC), and AMD's sold-out 2027 Venice (channel checks, not AMD).