← The Andrew Blog

Startup Finance / AI Economics

A 40% AI Compute Discount Can Cost You Five Months of Runway

The lower your cost per token, the better your margins. Unless you've committed to buying compute your customers never use.

Share on LinkedIn
Locked compute infrastructure beside declining cash reserves, illustrating the runway risk of minimum AI compute commitments.

In the first nine months of 2026, Anthropic reported a $104 billion GAAP net loss but generated $6.6 billion in operating cash flow. Its November 17 IPO filing also disclosed approximately $900.9 billion of aggregate commitments, substantially related to compute and infrastructure, across multiple years. The loss included major noncash financing adjustments; the commitments are not immediately payable and include conditional arrangements.

Three different financial measures, three different stories.

For a much smaller AI startup, the lesson is simple: A lower compute price can become an expensive decision when you commit to buying capacity you might never need.

The 40% Discount That Adds $45,000 to Monthly Burn

Consider a hypothetical seed-stage AI SaaS company:

MetricAmount
Monthly recurring revenue$100,000
Monthly variable inference expense$25,000
Compute-only contribution margin75%
Starting cash$1.2 million

The 75% contribution margin excludes other costs of revenue. It is not necessarily the company's GAAP gross margin.

A cloud vendor offers a 40% discount on compute in exchange for a $70,000 monthly minimum for twelve months.

Leadership is considering the offer because its growth forecast predicts substantially higher usage.

There are two choices.

Variable pricing: Pay the standard rate for actual compute consumed, with no minimum.

Committed pricing: Pay 60% of the standard rate, but at least $70,000 every month, regardless of consumption.

For this hypothetical contract, unused capacity expires monthly, additional usage receives the same discount, payments are made monthly, and there are no upfront fees or rollover credits.

The startup is guaranteeing the vendor $840,000 in annual spending. The vendor is not guaranteeing the startup enough customers to use it.

What the Commitment Does to Runway

Suppose the company spends another $160,000 monthly on payroll and operations, excluding inference.

If customer revenue remains at $100,000 per month, and customers pay in the same month, the cash picture changes dramatically.

Monthly cash economicsVariableCommitted
Customer receipts$100,000$100,000
Other operating payments$160,000$160,000
Compute payments$25,000$70,000
Net cash burn$85,000$130,000
Runway14.1 months9.2 months

A 40% compute discount just cost the company 4.9 months of runway.

This simplified downside calculation excludes taxes, capital expenditures and financing, and assumes receipts and operating payments remain constant. Under the commitment, the company runs out of cash before the twelve-month contract expires.

For a seed-stage startup preparing its next financing, those five months could determine whether it reaches its next major milestone.

There Are Two Break-Even Numbers, Not One

This is where many compute purchasing decisions go wrong.

At current demand, the startup spends $25,000 monthly at standard prices. Its committed minimum would be $70,000.

Threshold 1: Economic break-even

$70,000 ÷ $25,000 = 2.8× current demand

At 2.8 times current usage, both pricing options cost $70,000 per month.

Beyond this level, committed pricing starts saving money.

Threshold 2: Full utilization

$70,000 ÷ 0.60 = $116,667 of compute at standard prices.

$116,667 ÷ $25,000 = 4.67× current demand

That's the level at which the company fully consumes the discounted capacity included in its monthly minimum.

These thresholds mean different things.

At $75,000 of standard-price demand, the committed plan still costs $70,000. The startup saves $5,000 even though some discounted capacity goes unused.

You don't need full utilization to save money. But you do need sufficient demand to cover the minimum commitment.

Three Demand Forecasts, Three Financial Outcomes

Now examine what happens over twelve months.

Assume inference demand at standard prices equals 25% of customer revenue. All three scenarios use the same contract terms.

Annual resultsStrong growthFlat demandVolatile demand
Customer revenue$5.04M$1.20M$3.96M
Variable compute payments$1.260M$300K$990K
Committed compute payments$996K$840K$924K
Commitment savings+$264K−$540K+$66K
Variable contribution margin75.0%75.0%75.0%
Committed contribution margin80.2%30.0%76.7%

Hypothetical figures. The contribution margins cover revenue less inference expense only. For this example, compute expense and cash payments occur in the same month.

Strong growth: Monthly compute demand at standard prices increases by quarter from $25K to $75K, $140K and $180K. The commitment saves $264,000.

Flat demand: Usage stays at $25,000 monthly. The commitment costs $540,000 more than variable pricing.

Volatile demand: Monthly compute demand alternates by quarter between $25,000 and $140,000. The company still saves $66,000 compared with variable pricing.

But the volatile scenario reveals something the other two don't.

Average monthly compute demand is $82,500, above the $70,000 economic break-even.

If demand arrived evenly at $82,500 every month, the commitment would cost $840,000 annually versus $990,000 under variable pricing.

That's $150,000 in savings.

When demand alternates between $25,000 and $140,000, committed payments instead total $924,000.

Savings fall to $66,000.

The volatility alone costs $84,000 in potential savings, even though total annual demand is identical.

This distinction depends on the contract.

Our example assumes unused monthly capacity expires. Some enterprise agreements let eligible spending accumulate against an annual or multiyear commitment. Under full annual pooling, this particular volatility penalty would disappear.

Other arrangements are stricter. For example, AWS Savings Plans apply commitments hourly, without carrying unused amounts into future hours.

Before evaluating a discount, understand the period over which utilization is measured.

What If Better Models Make Compute Cheaper?

Customer demand isn't the only uncertainty.

Suppose the startup reaches a workload that costs $140,000 monthly at standard variable prices.

Under its commitment, it pays $84,000.

Now imagine two possible changes.

First, model efficiency improves. The same customer workload requires half as much compute.

Variable pricing falls to $70,000. Under the committed plan, discounted usage would theoretically cost $42,000, but the startup still pays its $70,000 minimum.

The discount no longer creates savings.

Second, an alternative provider cuts equivalent inference pricing by 50%. That workload would cost $70,000 at the alternative provider, versus $84,000 using the original committed provider.

Switching could be even more expensive because the existing minimum remains payable unless the contract permits termination or adjustment.

Neither scenario is a prediction. Both illustrate why infrastructure flexibility has financial value.

A commitment that looks attractive today can become a liability as model architectures, efficiency and provider pricing evolve.

Gross Margin Doesn't Show the Whole Exposure

Before signing, founders need three financial views.

Accounting profitability: What appears in revenue and cost of revenue, including the expenses recognized for delivering the service.

Cash economics: When customers pay, when infrastructure payments are due, and how those dates affect cash balances.

Contractual exposure: Minimum purchases, prepaid credits, termination costs and future spending obligations.

These are related, but they are not interchangeable.

For example, an upfront compute prepayment reduces cash immediately, while the related expense may be recognized as services are consumed.

A future minimum purchase commitment may be disclosed without the entire future amount becoming a current balance-sheet liability when the agreement is signed. Commenced leases can have different accounting treatment.

The founder's financial model needs to capture all three views.

The Compute Commitment Stress Test

Before committing to cloud or GPU capacity, run six checks.

1. Expected utilization

Forecast monthly demand based on customer activity, model usage and infrastructure efficiency. Don't rely solely on annual averages.

2. Minimum cash obligation

Calculate everything payable regardless of usage, including upfront payments, minimums and termination fees.

3. Break-even demand

For this hypothetical contract:

  • Variable monthly cost = V
  • Committed monthly cost = MAX($70,000, 0.60 × V)

Calculate both alternatives for every month. The commitment is economically advantageous over the contract term only if its aggregate payments are lower.

4. Downside cash exposure

Test demand 25%, 50% and 75% below plan. Include a temporary customer slowdown and a substantial improvement in model efficiency.

5. Runway impact

Project monthly ending cash balances under both alternatives. Identify the lowest cash balance and whether additional financing becomes necessary before the next milestone.

6. Contract flexibility

Determine whether unused credits roll forward, usage can move between models or regions, minimums can be reduced, prices can reset, and early termination is possible.

A longer commitment can be the correct decision when demand is demonstrably predictable, customers are diversified, and capacity availability protects contracted revenue.

Annual customer prepayments can also help, provided collection schedules and vendor payment obligations genuinely align.

But projected growth is not contracted demand.

And a cheaper unit price is not necessarily cheaper total spending.

The Decision Isn't Just About Price

Variable compute pricing lets a company reduce spending when demand changes.

A commitment trades some of that flexibility for lower prices and potentially more reliable capacity.

That can create real value. It can also consume the cash buffer a startup needs to survive a missed forecast.

The most important number in a compute agreement isn't the discount percentage. It's the minimum demand your startup must achieve before that discount creates value.

If an infrastructure commitment could materially change your runway or fundraising date, have an experienced fractional CFO stress-test the financial model and contract terms before signing.