As of late August 2026, Meta is not sitting on a warehouse of idle GPUs waiting to become “the next AWS for AI.” It is demand-constrained and overbuilding on purpose. Those two facts can be true at the same time. The interesting question is not “does Meta have spare compute?” but what kind of spare, when it appears, on which silicon, and whether it is commercially usable against OpenAI/Anthropic-style cloud.
The companion valuation article prices the same 1 GW under three uses: training at $0 direct revenue, $30 billion of assumed internal business value, and $50 billion of frontier-inference revenue. This deep dive supplies the physical denominator underneath that exercise: the campuses, silicon, workloads and capacity that could actually be exposed.
Prometheus in Ohio and Hyperion in Louisiana are the two flagship “Titan” campuses. Prometheus is already a multi-building, multi-hundred-MW operational cluster that includes rapid-deployment structures; Hyperion is the campus Meta has sized for as much as 5 GW of compute.
Table of contents
- Executive snapshot
- How big is the fleet, really? (GPUs, GW, sites)
- Hardware mix: NVIDIA, AMD, MTIA, CPUs
- Workload split: training vs product serving vs ads/recs
- Why Meta still buys from everyone else
- The Titan / Meta Compute buildout
- The “sell spare compute” thesis vs what Zuck actually said
- What capacity is structurally unused vs tactically unused
- Could this compete with OpenAI / Anthropic cloud?
- Podcasts, talks, and primary sources
- Working model of untapped capacity
- What would change the conclusion
1. Executive snapshot
The summary card above separates Meta's total build target from the much smaller merchant slice.
| Dimension | Best current picture (Aug 2026) |
|---|---|
| Self-described compute target | ~7 GW deployed in 2026, 14 GW by 2027 (internal memo via Reuters) |
| Tracked AI DC pipeline | ~15.8 GW across ~20 AI-tagged sites; only a slice is live today |
| GPU fleet (H100-eq) | ~1.3 million H100-equivalents by end of 2025, still ramping (Omar Baldonado / Hot Interconnects) |
| Largest live training-class cluster | 129k H100 cluster (2024–25), then Prometheus as a 1 GW+ multi-building system growing toward 3 GW+ |
| Custom silicon | Hundreds of thousands of MTIA chips already in production for ads/ranking inference; MTIA 300 in production for rec-sys training; 400/450/500 aimed at GenAI inference 2026–27 |
| External purchases | CoreWeave ~$21B expansion (through 2032), Nebius up to ~$27B, Crusoe ~1.6 GW, earlier Google ~$10B, AWS Graviton “tens of millions of cores,” Azure Foundry “hundreds of millions $/yr” in model tokens |
| Official stance on spare capacity | “We think we have a use for that compute.” Selling is “on the table” as a safety valve, not as evidence of idle GW today. Higher margin on selling intelligence than raw FLOPs. |
| Analyst “merchant” slice | 0.5–1.0 GW sold externally ≈ $11–22B/yr revenue in one Evercore estimate; that would be ~4–7% of a 14 GW 2027 fleet. |
The core tension: Meta is simultaneously the company building five 1 GW+ “Titan” campuses and one of the largest buyers of other people’s AI capacity. That is the tell. If there were large, clean, unused training clusters sitting idle, they would not still be signing multi-year, multi-billion contracts with CoreWeave, Nebius, Crusoe, AWS, and Google.
2. How big is the fleet, really?
Public numbers do not all measure the same thing. Distinguish four layers.
A. Accelerator count (Meta’s preferred unit: H100-equivalents) Hot Interconnects 2026: training clusters grew from ~24–32k → 129k → fleet 1.3 million H100-eq by end of 2025. Prometheus targeted online by end of 2026; Hyperion later at 5 GW. An earlier 2025 Hot Interconnects talk already put the GPU fleet above 600k H100-eq. The jump in one year is the story.
B. Power (IT MW vs campus GW) Epoch AI’s satellite-based tracker (updated 28 Aug 2026) puts Prometheus alone at ~562 MW IT / ~680k H100-eq now, heading toward ~1.0M H100-eq / ~785+ MW as more halls and on-site gas plants come up. Cheyenne and Kuna are each ~152 MW / ~173k H100-eq on B300-class silicon. These are IT numbers; facility power is 20–50% higher.
C. Named Titan campuses (SemiAnalysis) Five simultaneous 1 GW+ “Titans”: Prometheus (Ohio, expanding ~1 → >3 GW), Hyperion (Louisiana, 1.5 GW under construction now, designed for 5 GW), plus El Paso, Iowa (1 GW lease), Indiana. Prometheus is a constellation: 27 DCs across 6 campuses, five within 6 km, one 75–80 km away, stitched by an AI backbone SemiAnalysis puts at 22 Pb/s bidirectional.
D. Corporate pipeline vs live AI Data Center Index: 20 tracked Meta AI facilities, 15.8 GW selected capacity, ~50% operational by count, but operational MW is much smaller than nameplate (their own breakdown showed only ~2.4 GW operational in mid-June tracking — treat that as incomplete vs Meta’s 7 GW 2026 deployment claim). Hyperion is the single largest disclosed site at 5,000 MW. Gridwork permit data independently puts Meta at ~14 GW nameplate across nine U.S. states. Same order of magnitude as the Reuters 14 GW / 2027 target.
Capex context: 2026 guidance has been in the $115–145B range; one memo cited up to $145B this year. Future contractual commitments around cloud, servers, and DCs have been reported near $350B non-cancelable and almost $700B if you include broader long-term agreements. That is not the balance sheet of a company that already has spare capacity. It is the balance sheet of a company pre-paying the next three to five years of scarcity.
3. Hardware mix: not one cloud you can just rent
Meta runs a portfolio, not a single SKU. That matters for “can they sell this as cloud.”
NVIDIA Workhorse of Llama-scale pretraining. 24k H100 clusters (IB and RoCE) in 2023; 129k H100 cluster by emptying five production buildings; then Blackwell / GB200 “Catalina” racks (~72 GPUs, ~140 kW, ~360 PFLOPS FP16, air-assisted liquid cooling). Later GB300 / Rubin-class deployments via both self-build and CoreWeave. SemiAnalysis has been harsh on Meta’s custom “Ariel” GB200 config (1 B200 + 1 Grace instead of the standard 2+1), arguing it was worse TCO for LLM work; they say Meta is now buying the standard config.
AMD Feb 2026: long-term deal for up to 6 GW of Instinct GPUs, Helios rack-scale, shipments from H2 2026. Zuck framed it as diversification for efficient inference and “personal superintelligence.” Custom MI450X SKUs for Meta have been reported as cut-down (fewer XCDs / less I/O / lower HBM stack) — again, optimized for Meta’s mix, not for being a general merchant GPU.
MTIA (in-house) This is the part most cloud-thesis slides ignore.
- Already hundreds of thousands of MTIA chips in production, primarily ads and ranking inference.
- Four new generations in two years: 300 (rec-sys training, now in production, on-package 800G NICs, 1.2 TB/s I/O, 3.9× faster communication on a 150B rec model vs GPU), 400 (GenAI + R&R, 72-accelerator scale-up domain), 450 and 500 (inference-first, more HBM bandwidth/capacity). “Iris” was the internal name Reuters attached to a chip entering production around Sep 2026.
- Design philosophy is the inverse of NVIDIA’s: optimize first for Meta’s inference and rec-sys, then stretch into GenAI. That is excellent TCO for Facebook/Instagram. It is a terrible starting point for a public GPU cloud.
CPUs April 2026: multi-year, multi-billion AWS deal for tens of millions of Graviton5 cores — not accelerators, CPUs for agentic orchestration and CPU-side inference. Meta is renting Amazon silicon inside AWS data centers. That is the opposite of surplus AI accelerators.
Power side-bets Nuclear PPAs (Vistra PJM ~2.6 GW, Oklo Ohio, TerraPower Natrium), on-site / behind-the-meter gas (Socrates plants at Prometheus, Entergy gas around Hyperion), even a milestone-based deal for up to 1 GW of space-based solar later this decade. Power, not GPUs, is the multi-year constraint Zuck has been talking about since the 2024 Dwarkesh interview.
4. Workload split: training vs serving users vs internal products
Meta never publishes a clean pie chart. You have to reconstruct it from engineering posts and analyst models.
4.1 Three different computers inside one company
A. Ranking / recommendation / ads (the cash engine) This is still the majority of production AI at Meta, and a “meaningful portion” of total AI compute even in 2026.
- GEM (Generative Ads Recommendation Model) now trains at LLM scale on several thousand latest-gen GPUs; they report doubling E2E training efficiency to 20–25% MFU while scaling training FLOPs 4× in 12 months.
- Adaptive Ranking Model serves LLM-scale models in the ads path at ~100 ms, O(1T) parameters across multi-card serving, ~35% MFU, +3% ad conversions / +5% CTR on Instagram since Q4 2025.
- MTIA exists because rec-sys is communication-bound and embedding-heavy, not matmul-bound the way Llama pretraining is. One older industry note even claimed rec/ads-class models were the large majority of Meta’s historical AI FLOPs; that share is falling as GenAI grows, but it has not gone away.
This compute is not spare. It is load-bearing revenue. Every incremental ranking point is real ad dollars. You do not rent this out.
B. Frontier / Llama / Muse Spark training (Meta Superintelligence Labs) This is the bursty, high-glamour load.
- Llama 3 era: 16k–24k H100 clusters, ~90% effective training time, 38–43% MFU.
- Llama 4: training started on 32k GPUs and scaled toward ~100k in one synchronized job on the multi-building cluster.
- Post-training / RL is now a first-class consumer: LlamaRL, asynchronous RL across regions, thousands of engineers on RL environments (SemiAnalysis: ~3,000). Distributed training software (Twine, MAST) is being rewritten for multi-building and even multi-region jobs.
Training utilization profile: near-constant high power for weeks, then a hole when the run ends. That hole is the only structurally cyclical spare in the fleet. It is also the spare every lab wants to fill with the next run, with RL, or with evals.
SemiAnalysis’s important caveat: even after assigning a large slice of Meta compute to RecSys/ads, MSL training compute is still comparable to OpenAI and Anthropic through 2026–27, and their Tokenomics model has Meta ahead of both on total AI compute by end of 2026. “More compute than OpenAI” ≠ “idle compute.”
C. Product inference for users (Meta AI, Llama in products, agents, glasses, WhatsApp, etc.) Continuous, diurnal, latency-sensitive. Industry inference fleets often average 30–50% utilization against peak; Meta’s own ads-path MFU numbers (~35%) suggest they are better than average but nowhere near training-cluster saturation. As Meta AI, Muse Spark, coding agents, and on-platform assistants scale, this bucket grows and eats the training-cycle holes.
Zuck and finance have been explicit on earnings: they are demand-constrained, including in the core business — there are still ROI-positive places they would put compute if they had it.
4.2 A realistic split (order-of-magnitude, not official)
No source gives a 2026 official breakdown. A defensible working range:
| Bucket | Share of AI accelerators (working estimate) | Utilization character |
|---|---|---|
| Ads / feed ranking training + inference (MTIA + GPU) | 25–40% | High and continuous; not rentable |
| Llama / Muse / MSL pretrain + RL | 25–40% | Spiky; 70–90% during runs, much lower between them |
| Consumer + internal GenAI inference (Meta AI, agents, glasses, employee tools) | 15–30% | Diurnal; growing fast |
| Slack / evals / research / overflow | 5–15% | The only near-term merchant candidate |
Those ranges are reconstructed from SemiAnalysis (“meaningful portion” to RecSys), Meta engineering (MTIA “hundreds of thousands” already in ads inference), and the existence of 100k-GPU training jobs. Treat them as a map, not a census.
5. Why Meta still buys from everyone else (the challenge you asked for)
This is the strongest argument against “untapped GW ready to become a cloud.”
Confirmed or well-reported external capacity
| Counterparty | What was bought | Why it matters |
|---|---|---|
| CoreWeave | Expanded to ~$21B through Dec 2032; total relationship cited as high as ~$35B; includes early Vera Rubin | Dedicated GPU capacity because self-build lagged demand. Meta is CoreWeave’s anchor. If Meta becomes a seller, CoreWeave’s biggest customer becomes a competitor. |
| Nebius | Up to ~$27B | Same pattern: lock scarce accelerators years ahead. Combined CoreWeave+Nebius ~$48B is the figure that spooked neocloud stocks in July. |
| Crusoe | ~1.6 GW across Texas and Missouri | Power + building, not just chips. Meta is buying time. |
| Google Cloud | ~$10B / 6-year (2025) | Raw infra scaling across all three hyperscalers. |
| AWS | Multi-year, billions, tens of millions of Graviton5 cores | CPU side of agentic stacks. Meta is a tenant of a direct competitor. |
| Microsoft Foundry | Hundreds of millions $/year; trillions of tokens per week | Meta is buying other labs’ models (including OpenAI via Azure) to benchmark and to run production tools while its own models catch up. Simultaneously building a rival API. This is the Bing-then-replace pattern. |
| Colocation / leases | Iowa 1 GW lease; 5 GW+ of cloud+colo contracted in H1 2026 alone per SemiAnalysis | Self-build is not fast enough even with tents. |
What this implies
- 2024–2026 Meta was a net importer of AI compute. The 129k cluster was created by emptying five production DCs — they cannibalized existing estate because new build was too slow.
- Tents, off-grid gas, and neocloud contracts are the same strategy: collapse time-to-FLOPs. Distilled.earth called the tent + jet-turbine phase “Mad Max.” That is not how you behave when you have spare capacity. It is how you behave when every week of delay is a model-generation missed.
- Buying models on Azure while building Muse Spark means even software capacity is imported. Trillions of tokens/week is not a rounding error; it is Meta using someone else’s inference fleet as a bridge.
- The $48B neocloud question is two-sided. Bears say Meta will walk away and dump spare capacity onto the market. SemiAnalysis’s counter is that Meta contracted another 5 GW+ in H1 2026 and 2027 capex will still be “shockingly high.” Both can be true in sequence: import now, maybe export a thin slice in 2027–28 if Titans land faster than MSL and ads can absorb them.
- Sceptics noted Meta could not even get the Google capacity it wanted. That is the opposite of a glut.
An internal-facing talk, “Extending Meta to the Public Cloud,” is unusually honest: they had CPUs and storage; GPUs were saturated; they could not build GPU DCs overnight; they onboarded a cloud partner in <3 months because every idle day was wasted money and a delayed model. Cloud is described as a permanent, growing part of Meta infra, with new vendors onboardable in ~a month. That is a buyer’s operating system, not a seller’s.
6. The Titan / Meta Compute buildout
January 2026: Zuck launches Meta Compute — “tens of gigawatts this decade, hundreds of gigawatts or more over time.” Santosh Janardhan on architecture/fleet; Daniel Gross on long-term capacity strategy; Dina Powell McCormick on governments and financing.
What “Titan” actually means in engineering terms (Omar Baldonado, Hot Interconnects 2026):
- Scale-up: inside a rack / NVL-class domain (512–576 accelerators today; they want >1,000 packages later).
- Scale-out: building / region Ethernet fabrics (they moved off InfiniBand after the 32k Llama 3 clusters).
- Scale-across: multi-building and multi-region, 1–10 µs inside a region, >500 µs beyond 100 km. Prometheus already spans tens of km; they talk about campuses 2,000 km apart for asynchronous RL.
Prometheus is the prototype of “one logical 1–3 GW cluster that is not one building.” Hyperion is the prototype of “one campus that is a small city’s worth of power.” That architecture is world-class for Meta’s training. It is awkward to productize as a public cloud region.
7. The “sell spare compute” thesis vs what Zuck actually said
Timeline of the rumor vs the words
- May 2026 shareholder meeting: cloud is “definitely on the table.” Companies ask “almost every week” for an API or for compute at a premium. They have not done it because “we think we have a use for the compute.” If they overbuild, it becomes an option — and that optionality is part of why they are willing to spend.
- July 2026 Bloomberg: Meta is building a plan for a cloud business — hosted models (Bedrock-like, including Muse Spark) and/or bare metal to neoclouds. Stocks: Meta up, CoreWeave/Nebius down 10–17%.
- July earnings / interviews: offers arrive at a significant premium to what Meta paid. Selling all of it would be “foolish” because margin on intelligence > margin on raw compute. Anthropic was even reported in preliminary talks to lease from Meta. Cloud is a safety valve, not a declaration of surplus. “I don’t think anyone in the industry feels like they have too much compute.”
Two products that are easy to confuse
- Hosted intelligence (API / Muse Spark / Llama as a service). This is the product Zuck prefers. It uses their software stack, their models, their MTIA+GPU mix. Competes with OpenAI API and Anthropic API more than with AWS EC2.
- Merchant FLOPs / bare metal. This is what would actually hit CoreWeave/Nebius. Harder: isolation, SLAs, NVIDIA/AMD licensing, CUDA vs ROCm vs MTIA, claw-back rights (one Stratechery-style script even imagined short-term rentals Meta can yank back). Evercore’s 0.5–1 GW “thin slice” is this bucket.
Zuck is consistently louder on (1) than on (2).
8. What capacity is structurally unused vs tactically unused
Structurally hard to monetize
- MTIA fleet. Wrong ISA, wrong networking story, wrong customer. Valuable only inside Meta’s rec/ads/GenAI-inference stack.
- Ads/ranking clusters. Revenue-critical, latency-critical, data-gravity to Meta’s user graph. Not leaving the building.
- Heterogeneous SKUs. Meta engineering has said that 5–6 new SKUs a year make the fleet hard to schedule and leave hardware underutilized internally. If they cannot move their own jobs cleanly, they cannot sell a clean instance type.
- Custom racks (Ariel, cut-down MI450X). Optimized (sometimes badly) for Meta, not for a price list.
Tactically unused (the real pool)
- Inter-run gaps on training Titans. After Llama 4 / Muse / Behemoth-class runs, a 50–100k GPU island can drop from ~90% busy to “only” inference, evals, and smaller experiments. That is the classic 30–50% post-training utilization cliff.
- New halls that energize before the model team is ready. Tents cut build time in half; software, data, and RL environments do not always keep up. Short windows of “commissioned but not yet swallowed by MSL.”
- Off-peak inference. Diurnal troughs on serving fleets. Hard to sell as reserved training capacity; possible as spot/preemptible.
- Older H100/H200 as newer Blackwell/Rubin/MTIA-400 land. The waterfall into inference is real industry-wide; Meta can either soak it internally or list it.
What is not unused
Almost everything Meta is paying CoreWeave, Nebius, Crusoe, Google, and AWS for. Those contracts exist because the internal pool was already spoken for.
A useful analogy: Meta is building a private power company that still buys from the grid at peak. The existence of a 5 GW plant under construction does not mean 5 GW is for sale this quarter.
9. Could this compete with OpenAI / Anthropic cloud?
Depends which market.
Against OpenAI / Anthropic APIs (intelligence): Yes, in principle, and this is the fight Meta actually wants. Distribution (3B+ users), open-weight Llama flywheel, and a coming first-party API. Quality is the constraint, not FLOPs. SemiAnalysis’s July 2026 progress note: Muse Spark still around Opus 4.6 level — lots of compute, not yet a quality lead. Compute surplus does not automatically become API share.
Against AWS/Azure/GCP general cloud: No, not in the next 24 months in any serious way. No IAM heritage, no enterprise sales machine, no global regions productized for tenants, no compliance catalog. Zuck has never run this business.
Against CoreWeave / Nebius / Lambda (neocloud GPU): Only as a thin, claw-backable, bursty seller. 0.5–1 GW at $11–22B/yr is material to those companies’ valuations (hence the July drawdown) and immaterial to Meta’s 14 GW plan. The moment Meta needs the GPUs for the next pretrain, the merchant book gets yanked. Sophisticated tenants know that.
Against “Anthropic leases Meta GPUs”: That rumor is the most interesting variant: Meta as wholesale to a lab that can pay a premium and tolerate weaker SLAs than a Fortune 500. That uses the inter-run gap without Meta having to become AWS. It also confirms there is no giant unused pile — if there were, they would already be filling it with MSL jobs.
Barriers that keep “untapped” from becoming “product”:
- Multi-tenant isolation on fabrics designed for one tenant (Meta).
- CUDA software moat vs MTIA/AMD mix.
- Claw-back vs customer trust.
- Sales/support culture.
- Cannibalizing suppliers Meta still needs through 2032.
- Zuck’s own preference: sell the model, keep the FLOPs.
10. Podcasts, talks, and primary sources worth listening to
| Source | Why it matters |
|---|---|
| Dwarkesh Patel × Zuckerberg (Apr 2024) | Best early primary on energy as the bottleneck, “no one has built a 1 GW training cluster yet,” 24k clusters, whole-fleet GPU counts. The mental model for everything that followed. |
| Dwarkesh Patel × Zuckerberg (2025, “AI will write most Meta code…”) | Physical infra lead times, NVIDIA generation lag, permitting, China industrial policy. Explicit rejection of pure software fast-takeoff. |
| Omar Baldonado, IEEE Hot Interconnects 2025 and 2026 | The best engineering primary: 129k → 1.3M H100-eq, Ethernet vs IB, scale-up/out/across, Prometheus/Hyperion numbers, zettaflop-era networking. |
| Engineering at Meta, Sep 2025, “Infrastructure Evolution and the Advent of AI” | Official history: 4k → 24k → 129k, emptying five DCs, Catalina racks, Prometheus 1 GW, Hyperion 5 GW, MTIA for ads, Twine/MAST for geo-distributed training. |
| Engineering at Meta, Aug 2026, MTIA 300 | Why rec-sys training is a different computer from LLM training. On-package NICs, HCCL, 3.9× comms. |
| “Extending Meta to the Public Cloud” (Govindasamy & Dhillon, 2026) | How and why Meta became a cloud customer; onboarding in weeks; GPUs were the scarce resource. |
| Pivot / Access (Zuck on the AI bubble and glasses) | Capex philosophy, “glimpses” of self-improving systems, political economy of the buildout. |
| Bloomberg exclusive with Zuck (Jul 2026) and Q2 2026 earnings | Cloud “on the table,” premium inbound offers, intelligence > compute margins, demand-constrained. |
| SemiAnalysis: “Future of Meta Superintelligence” (Jul 2026) and “Infra team needs a culture reset” (Jul 2026) | Best outside model of Titans, Tokenomics vs OpenAI/Anthropic, and the uncomfortable view that Meta infra has also shipped some bad custom SKUs. |
There is no single podcast that answers “how many spare GW.” The closest composite is Baldonado (what exists) + Zuck/Dwarkesh (why they will not sell it all) + the cloud-onboarding talk (why they still buy).
11. Working model of untapped capacity
Put it in one paragraph:
Today (Aug 2026): effectively ~0 GW of committed, multi-month, clean training capacity sitting idle. The fleet is demand-constrained. Apparent slack is either (a) MTIA/ads silicon that cannot be sold, (b) diurnal inference troughs, or (c) days-to-weeks between training runs. Meta is a large importer of GPU, CPU, and even rival-model tokens.
2027, if Titans land on schedule and MSL quality does not suddenly require every new rack: a tactical merchant slice of 0.5–1.5 GW is plausible — Evercore’s 0.5–1 GW is the base case, 1.5 GW if Hyperion ramps faster than product inference. That is 4–10% of a 14 GW stack. Revenue interesting (~$10–30B/yr at scarcity prices), strategy optional, not a new hyperscaler.
Structural unused over a full cycle: maybe 15–30% of GPU (not MTIA) nameplate averaged across the year, if you count inter-run gaps and inference diurnal patterns. Most of that will be consumed by RL, evals, internal agents, and Meta AI growth long before it is productized. Jevons applies inside Meta as much as outside: cheaper/more FLOPs get eaten by longer RL, bigger ads models, and more user-facing inference.
The capacity that looks “untapped” from a satellite is often already contracted internally or externally. Nameplate GW ≠ available reserved instances.
12. What would change the conclusion
Watch these, not the press releases:
- Do CoreWeave / Nebius / Crusoe renewals shrink in 2027 filings? If Meta cuts imported GW while Titans light up, that is the first real surplus signal.
- Does Meta publish instance types, an API catalog, or a reserved-capacity SKU? Words on an earnings call are optionality; a price list is a business.
- Muse Spark / Llama quality vs Opus / GPT. If they pull ahead, they will keep every FLOP. If they plateau, finance will push merchant compute.
- MTIA 400/450 actual deployment mix. If GenAI inference moves off NVIDIA onto MTIA at scale, a lot of H100/B200 capacity is suddenly “free” — that is the most likely large unlock, and it still may be kept for Llama.
- Power, not chips. If Hyperion gas/nuclear slips, the 14 GW 2027 number is paper. Spare capacity cannot exist on a campus that is not energized.
- Utilization leaks. If reliable reporting ever shows Meta GPU MFU collapsing toward xAI-like teens instead of the ~40%+ they have historically run, that is either incompetence or genuine idle. Current public crumbs (Llama 3 MFU, GEM 20–25%, Adaptive Ranking 35%, one comparison putting Meta ~43% vs xAI 11%) do not look like a glut.
How this connects to the Meta model
Our Meta model already assigns the infrastructure build to advertising, Business AI and consumer Meta AI. It does not treat 14 GW as a separate cloud-revenue line. Doing so would double count the internal value before Meta has disclosed a customer, a reserved MW figure or a price.
The companion 1 GW sensitivity keeps the distinction explicit: internal training adds no direct revenue; $30 million per MW-year is an internal-value assumption; and $50 million is the external frontier-inference case based on an analyst estimate of Anthropic's peak monetization. The published base model remains $757.11 per share until an external product becomes incremental and measurable.
Bottom line
Meta is building one of the two or three largest AI estates on earth, with a custom-silicon program that already serves the ads machine at huge scale and a training fabric that has crossed the 100k-GPU and 1 GW thresholds. The untapped part is real but narrow: cyclical holes between frontier runs, older GPUs as new ones arrive, and a political option to sell a thin slice so Wall Street stops treating $100B+ capex as pure incineration. It is not a hidden OpenAI-scale cloud waiting to be switched on. The company that is still writing $20B+ checks to CoreWeave and burning rival tokens on Azure is not the company with a secret surplus. The surplus, if it comes, will be a 2027–28 management choice about whether intelligence or kilowatt-hours pays more — and Zuck has already said which one he prefers.
Sources and provenance. Meta's infrastructure history, cluster progression and workload design come from Meta's infrastructure engineering review and its MTIA 300 disclosure. The $21B CoreWeave expansion and Nebius agreement worth up to $27B are company disclosures. Meta's stance on selling compute comes from its 2026 shareholder meeting and second-quarter call. Reuters, Epoch AI, SemiAnalysis and the talks identified in section 10 supply the reported fleet, campus and analyst estimates. The 0.5–1.5 GW merchant range, 15–30% structural-slack range and every conclusion drawn from them are scenarios, not Meta guidance.
