news

Meta's 14 GW Is Not One Revenue Pool. One GW Spans $0 to a $50B Analyst Upside Case.

Meta's 14 GW is not one revenue pool. We value 1 GW three ways: $0 direct revenue in training, $30B of internal business value, or $50B from frontier inference.

The same 1 GW spans $0 revenue to $50B

Annual economics · 1,000 MW at full effective use

1 GW operating modeAnnual economics
Internal training$0 direct revenue
Compute cost$10–15B
Internal Meta value$30B assumed
External raw compute$10–25B revenue
Frontier inference$50B revenue
Meta 2027 fleet share7.1%

Training revenue is $0 by classification. $30B is an R40 internal-value assumption, not Meta revenue. $10–25B is the external compute/service range. $50B applies Dylan Patel's analyst estimate of Anthropic's peak inference revenue per MW; it is neither an Anthropic price list nor a raw-compute rate.

One gigawatt is not a revenue number. It is an input whose value depends entirely on what Meta runs on it.

The same 1,000 MW can sit in three very different economic states:

That creates a $0 / $30B / $50B annual ladder for 1 GW. The ladder is useful only if its labels remain attached. Training is cost. Internal use is implied economic value. The $50 billion case is product revenue from paid inference at full load—not money generated merely by owning a cluster.

The full infrastructure deep dive maps Meta's Titan campuses, NVIDIA/AMD/MTIA hardware, workload split, outside purchases and the narrow capacity pool that could realistically become merchant compute. This article prices that pool rather than repeating the inventory.

Executive snapshot

1 GW operating mode Annual economics Classification
Internal model training $0 direct revenue Cost / future option
Internal ads, engagement and agents $30B implied value R40 assumption
External raw GPU service $10–25B revenue Market range
External frontier inference $50B revenue Analyst high-end anchor
Compute cost $10–15B cost Analyst range
Share of Meta's stated 2027 fleet 7.1% 1 GW ÷ 14 GW

All revenue and value rows assume 1,000 billable or productively deployed MW for a full year. At 75% effective utilization, the $30 billion and $50 billion cases become $22.5 billion and $37.5 billion. Neither should be multiplied across Meta's entire 14 GW fleet: much of that capacity is still being built, has non-billable facility load, or is already committed to core workloads.

The evidence behind the $10M, $30M and $50M layers

$10–15M per MW-year: the compute cost

On the 25 August 2026 Dwarkesh episode, SemiAnalysis founder Dylan Patel described the base annual cost of compute as roughly $10 million, $13 million or $15 million per MW, depending on the system. That is an analyst estimate, and it brackets the $12.0 billion per GW-year infrastructure proxy we build independently below.

This is the right baseline for internal training. A training run consumes a roughly $10–15 million-per-MW annual resource and reports zero direct revenue during the run. Its return arrives later, if the model makes another product more valuable.

$30M per MW-year: Meta's internal value hurdle

Meta does not disclose value per internal AI MW. We therefore use $30 million per MW-year as an explicit R40 midpoint assumption—not as revenue and not as guidance.

The number sits above the $10–15 million compute-cost range because keeping the MW only makes sense if Meta expects more value from ads, recommendations, agents and engagement than it could obtain by renting the hardware. It also sits below the observed $50 million frontier-inference anchor because Meta has not yet demonstrated Anthropic-like paid demand for its public model products.

At 1 GW, the assumption means $30 billion of annual business value. Against the analyst cost range, that is 2.0–3.0× gross value-to-compute cost. It is not $30 billion that should be added to Meta's reported revenue: most of the value would appear inside advertising or existing product lines, where the model already captures it.

$50M per MW-year: selling intelligence at the frontier

Patel said Anthropic's revenue from serving its frontier models had gone as high as $50 million per MW, while the underlying inference capacity cost about $10 million. Elsewhere in the same conversation he used $60-plus billion per GW as a talking number and described current monetization around $30–40 million per MW with a possible $60–70 million future range.

That makes $50 million per MW-year the cleaner base for the upside case. It is the observed-peak analyst estimate; $60 million is the round-number extrapolation. Neither is an Anthropic filing, customer contract or advertised SKU. The customer buys Claude tokens, seats and coding output—not a megawatt.

At full utilization, the translation is simple:

1,000 MW × $50M per MW-year = $50B of annual frontier-inference revenue.

Against $10–15 billion of annual compute cost, that is 3.3–5.0× revenue-to-compute cost before research, sales, model development, idle capacity and corporate overhead. At 75% effective utilization it is $37.5 billion, or 2.5–3.75× the same cost range.

Why $50M is not the price of raw compute

Raw infrastructure and finished intelligence occupy different layers of the stack.

Rackspace told investors that it expects $15–20 million of annual revenue per deployed MW as its GPU and customer mix evolves. It translated 30 MW into $450–600 million of annual revenue and anticipated adjusted EBITDA margins above 50%. Rackspace explicitly labels those figures illustrative and dependent on GPU model, deployment mix and customer terms.

IREN reported recent three-year contracts above $20 million per IT MW-year, with active discussions near $25 million. Its annualized measure multiplies contracted GPU-hour pricing by 8,760 hours and includes storage and ancillary services. IREN warns that recognized GAAP revenue can be materially lower.

Those disclosures establish a usable $15–25 million external AI-cloud band, inside Patel's broader $10–25 million raw-compute framing:

Public anchor Annualized revenue per MW What it represents
Rackspace low case $15M Illustrative enterprise AI
Rackspace high case $20M Illustrative enterprise AI
IREN recent contracts Above $20M Contracted AI cloud
IREN active discussions About $25M Prospective AI cloud

The evidence supports $15 million as a conservative external-service price and $25 million as the current upper public-company anchor. It does not support applying either rate to a cluster occupied by Meta's own training jobs, or calling $50 million the rent for GPUs.

The other source of confusion is Anthropic's reported contract with SpaceX. The SpaceX European prospectus discloses $1.25 billion per month through May 2029 for services across Colossus and Colossus II, with either party able to terminate on 90 days' notice after the initial period. That is $15 billion per year, but the document does not give a clean MW denominator for the contracted scope.

Using different illustrative denominators for the covered power produces very different rates:

Assumed contracted capacity $15B annual payment per MW
250 MW $60M/MW-year
300 MW $50M/MW-year
450 MW $33M/MW-year

The prospectus discloses approximately 325,000 GPUs but no contracted MW figure, so none of those three denominators is official. The arithmetic makes $60 million plausible under a tight denominator, not disclosed. Colossus is also a live, GPU-filled, short-notice cluster—not a 20-year powered shell. Anthropic's longer-term campus contracts show how large the product difference is: its 401 MW TeraWulf lease works out to $2.37 million per MW-year, while reported six-year Rubin compute deals are around $15–16 million.

It is possible for an inference product to create more value than GPU rental. But once a company sells API tokens, subscriptions, agents or business outcomes, MW is only an infrastructure input. The revenue equation becomes:

Revenue = delivered tokens × price per token × paid demand

The capacity equation underneath it is:

Delivered tokens = MW × tokens per watt × utilization × time

Every term moves. Model architecture changes token throughput. Quantization and serving software change tokens per watt. Input and output tokens carry different prices. Long reasoning chains consume more compute. Batch size and latency targets change utilization. A free assistant can use the entire cluster and still report no direct revenue.

NVIDIA's own 2026 analysis illustrates the dispersion. It estimates 54,000 tokens per second per MW on H200 and 2.8 million on GB300 for one benchmarked workload—a greater than 50-fold difference. That makes a fixed revenue-per-MW assumption unstable across hardware and models. NVIDIA also presents vendor scenarios in which token revenue far exceeds hardware cost, but those require sustained paid demand at the assumed token price. They are not customer contracts and do not validate $60 million per MW-year for Meta.

The verified treatment is therefore:

What 1 GW of internal training costs

Meta has not disclosed a complete cost per GW, so the following is an R40 economic proxy, not a company figure.

The physical inputs are $9 million of site infrastructure per MW, $65 per MWh of electricity, 1.2 PUE and annual site operating costs equal to 8% of site capex, taken from a 2026 hyperscale design model. For GPUs and related systems, we use $40 million per MW as a conservative upper bound. IREN defines payback as GPU-and-ancillary capex divided by contracted revenue less estimated direct costs, and says recent contracts above $20 million per MW pay back in around two years. Setting direct costs to zero produces the highest implied GPU capex: roughly $20 million multiplied by two. The actual figure should be lower because IREN's denominator is revenue after direct costs. Site infrastructure is separate, so we do not subtract the $9 million site assumption from this GPU capex bound.

We recover the site over 15 years and the GPU-and-systems layer over four. That is an economic replacement allowance, not Meta's accounting policy.

1 GW annual cost proxy Full load 75% compute use
Electricity $0.683B $0.512B
Fixed site operations $0.720B $0.720B
Facility cash-cost proxy $1.403B $1.232B
Site capital recovery $0.600B $0.600B
GPU and systems recovery $10.000B $10.000B
Economic infrastructure cost $12.003B $11.832B
Direct external training revenue $0 $0

This is still incomplete. It excludes model researchers, data acquisition and cleaning, failed runs, checkpoint storage, software engineering, security and the opportunity cost of keeping the cluster away from ads or an external customer.

Training can generate an excellent return. It just does so indirectly: a stronger recommendation model can lift ad conversion; a frontier model can support a paid API; an agent can create a new business. Until one of those outcomes is measured, the training cluster belongs on the cost side.

The three economic states of the same capacity

1. Train an internal model

Meta keeps the entire output: weights, research knowledge and the option to deploy the model across its products. The cluster produces no external revenue while the run is underway. The investment earns only if the resulting intelligence improves an existing business or creates a new one.

This is the use Zuckerberg prefers because successful intelligence can compound across billions of users. It is also the hardest return to isolate: ads, engagement and product revenue can improve without a separate “training revenue” line ever appearing.

2. Deploy it inside Meta

Meta can run the trained intelligence across recommendations, advertising, WhatsApp agents, creator tools and employee workflows. This is where our $30 million per MW-year assumption belongs.

It is a hurdle rate for economic value, not a new business segment. If an AI recommendation model creates $30 billion of incremental ad value on 1 GW, the benefit appears as better pricing, conversion or engagement inside Family of Apps. Counting another $30 billion of “compute revenue” would count the same economics twice.

3. Sell outside Meta

Meta sells a reservation, GPU-hours or an interruptible block to an outside lab. This is where the $10–25 million per MW-year compute-service range is relevant.

The product is easier to measure and finance. A take-or-pay customer contract supports hardware procurement; a spot window between Meta training runs does not. Revenue depends on how much of the 1 GW is genuinely billable and for how long.

The higher-value version is to keep control of the compute and sell tokens, subscriptions, agents or outcomes. That is the $50 million per MW-year Anthropic-level case. The revenue can exceed raw rental because the customer is buying intelligence rather than electricity or a GPU.

This category still requires a bottom-up model of paid users, token volumes, prices and serving efficiency. The $50 million rate describes what a loaded, well-monetized inference fleet may produce. It does not turn every idle Meta megawatt into $50 million automatically.

One gigawatt is not one stable quantity

Even the cost denominator needs care.

That is why this article uses MW-year only for external contracts and uses cost for internal training. Converting all nameplate GW into revenue would erase the distinction management is actually deciding between.

What capacity can realistically be exposed

Meta's estate contains several workload pools.

Ranking, recommendation and ads run continuously and support the cash engine. Meta's MTIA chips are designed around these workloads. This capacity is difficult to sell because it is tied to Meta data and can already produce returns inside the ad auction.

Frontier training and reinforcement learning create the most visible gaps. A large synchronized run can occupy a cluster for weeks and then end. The gap before the next run is real capacity, but it is suitable mainly for interruptible customers unless Meta makes a permanent allocation.

Consumer and enterprise inference is continuous and diurnal. Off-peak capacity can support a spot service; reserved annual contracts require Meta to hold enough capacity away from its own peak demand.

Older NVIDIA accelerators are the cleanest merchant pool as new generations arrive. They have the software ecosystem outside customers expect. Moving internal inference onto MTIA could release them without creating a company-wide surplus.

The $48 billion evidence that Meta is still a buyer

The strongest evidence against a hidden, clean 1 GW surplus is what Meta continues to purchase.

CoreWeave disclosed a $21 billion expansion through December 2032, including early NVIDIA Vera Rubin deployments. Nebius disclosed an agreement worth up to $27 billion: $12 billion of dedicated capacity beginning in early 2027 plus up to $15 billion of additional capacity that Nebius may first sell elsewhere.

Those announced maximums sum to $48 billion. Buying and selling can coexist—Meta may need new NVIDIA systems in one region while an older cohort or a temporary training gap is available in another. The contracts do rule out the simple thesis that a homogeneous 1 GW block is sitting idle and ready for a multi-year external customer today.

What management has and has not promised

At the May shareholder meeting, Zuckerberg said a cloud business was "definitely on the table" and that companies approached Meta almost every week for an API or compute at a premium. He also said Meta had not done it because it believed it had a use for the compute; selling was an option if it overbuilt.

By the July earnings call, Meta had launched a public model API. Zuckerberg described direct compute sales as an opportunity but said there was nowhere near enough compute for total demand and that selling intelligence should have a significantly higher margin. CFO Susan Li said Meta remained supply constrained even in its core business.

Management has disclosed inbound demand. It has not disclosed an externally reserved MW figure, customer contract, utilization, price, hardware cohort or direct compute revenue.

What the three assumptions do to our Meta model

The published Meta model carries $757.11 per share in its base case. It already charges the infrastructure build primarily to advertising and contains separate future Business AI and consumer Meta AI lines.

For training-only capacity, the correction adds no revenue and no new vertical. Its cost and possible returns already live in the model's capex, advertising and AI assumptions. The model therefore remains at $757.11.

The $30 billion internal-value case is not added. Our model already puts AI benefits into advertising, Business AI and consumer Meta AI; adding a second “internal compute” line would double count them.

We did run a conditional external sensitivity—not a model change. It assumes Meta monetizes a full 1,000 MW beginning in 2027 Q1, earns a 20% incremental free-cash-flow margin, and retains the model's 9% discount rate, 6x terminal revenue multiple and 2.566 billion diluted shares. We apply the same margin to every row so the rate, rather than a hidden margin change, drives the result.

Conditional external service Fair value Change
No customer; internal training only $757.11 $0
$15M/MW-year, fully contracted $784.04 +$26.93
$20M/MW-year, fully contracted $793.01 +$35.90
$25M/MW-year, fully contracted $801.99 +$44.88
$50M/MW-year frontier inference $846.86 +$89.75
Lower the base exit multiple to 5x $637.90 −$119.21
Raise the base exit multiple to 7x $876.32 +$119.21

The $15–25 million rows show a separately contracted infrastructure service. The $50 million row shows a different product: Meta achieving Anthropic-like inference revenue and paid utilization. It is an upside test, not a forecast, and it overlaps the model's existing AI verticals unless Meta creates a genuinely incremental external product. No customer or paid inference means no external revenue.

The comparison also keeps the valuation honest. Even the $50 million frontier-product case moves fair value less than changing Meta's terminal multiple by one turn. The ordinary valuation assumption remains more important than the hypothetical cloud business.

What to watch

  1. A signed external contract. Inbound offers are not revenue.
  2. Reserved IT capacity. Meta needs to disclose MW or accelerators and the contract duration.
  3. Training versus inference. Training is cost; inference can be monetized per token or user.
  4. Recognized revenue versus ARR. IREN explicitly warns the two can differ materially.
  5. Utilization and claw-back rights. Temporary gaps cannot support the same economics as firm annual capacity.
  6. MTIA migration. Moving internal workloads off NVIDIA is the cleanest route to rentable GPUs.
  7. A paid product metric. Users, tokens, price or agent outcomes are required before assigning end-customer revenue to a megawatt.

Bottom line

One gigawatt of Meta training is a large infrastructure and research expense. In our cost proxy it requires about $1.4 billion of annual facility cash cost and $12.0 billion of economic infrastructure cost, before the researchers and data. Its direct external revenue is zero.

Our middle case assigns $30 billion of annual internal business value to that 1 GW. That is an R40 assumption for the benefit to ads, recommendations and products—not revenue Meta reports and not a number to add on top of those businesses.

If Meta sells raw capacity, the evidence points to roughly $10–25 billion per GW-year. If it matches Anthropic's analyst-estimated peak economics by selling frontier intelligence, $50 billion per GW-year is the better upside anchor. The $60 billion shorthand is plausible as a high case, but it is neither an official Anthropic price nor the rent for an ordinary megawatt.

That is the answer to the headline. Meta has vast compute, but not uncapped, homogeneous or idle compute. The revenue opportunity exists only when Meta moves a usable slice away from training, finds paid demand and sells the intelligence at a higher value than the work it displaced.


Meta's fleet targets are from a July 2026 internal memo reported by Reuters; workload and silicon details come from Meta engineering and its MTIA 300 disclosure. External-service rates come from Rackspace's illustrative outlook and IREN's contracted ARR disclosure; Dylan Patel's Dwarkesh interview supplies the $10–15M compute-cost and $50M peak Anthropic-revenue estimates; SpaceX supplies the $15B annual Colossus payment but not a definitive MW denominator; NVIDIA supplies the hardware-specific inference-throughput comparison. The $30M internal-value rate and all 1 GW cost, utilization, capital-recovery, financing and fair-value figures are R40 assumptions and arithmetic, not Meta guidance. No direct revenue is assigned to internal training, and the $50M rate is inference-product revenue rather than an Anthropic MW SKU.

Related

Stocks in this article