A seed-stage AI startup generates $100,000 in monthly revenue and spends $20,000 on inference.
Its compute-only gross margin is 80%.
Then its cloud provider offers a deal: commit $400,000 upfront and receive a 25% discount on compute.
The founder signs.
In the first month, recognized compute expense drops from $20,000 to $15,000. The compute-only gross margin improves to 85%.
But what happens if customer growth doesn't follow the pitch deck?
Twelve months later, the company has spent $400,000 in cash on compute while generating $1.2 million in revenue.
That's an effective cash-based compute margin of just 67%.
The accounting records may have shown an 85% margin on consumed compute before the unused credits were written off. Whether that write-off affects reported gross margin depends on its accounting classification.
Either way, the economics are worse than the original 80% pay-as-you-go margin.
And the company has $160,000 less cash than it would have had without the commitment.
The dangerous decision wasn't paying too much for compute.
It was committing cash against growth that hadn't happened yet.
This is a hypothetical startup. The economics are illustrative, but the contractual risks are real.
When the margin improves but cash economics get worse
Reported compute-only gross margin on consumed credits, before unused-credit write-offs.
Effective cash-based compute margin when unused credits expire.
$400,000 paid upfront. $180,000 of credits consumed. $220,000 of unused credits lost.
The $3.2 Trillion AI Infrastructure Story Has a Smaller Lesson
In August 2026, credit rating agency KBRA reported approximately $3.2 trillion in gross disclosed contractual commitments and maximum contingent support across seven major AI infrastructure companies.
That compared with approximately $575 billion at the end of 2024.
The figure includes leases, purchase and construction commitments, supply arrangements, cloud capacity agreements, and guarantees.
It is not $3.2 trillion of debt or immediately payable cash obligations. Fixed obligations and contingent exposures carry different risks.
Microsoft's September 2026 quarterly filing disclosed $364.6 billion in additional leases not yet commenced, primarily for data centers. Some arrangements were subject to contractual conditions.
Oracle disclosed $288 billion in additional lease commitments as of August 31, 2026, plus $34.15 billion in unconditional purchase and other obligations. These represent different contractual categories, generally with payments scheduled over future years.
Sources: Microsoft Form 10-Q | Oracle Form 10-Q
But one detail in Oracle's October 2026 investor presentation deserves more attention.
Oracle identified customer prepayments as a way to fund hardware procurement before deployment.
Source: Oracle Investor Day, CFO presentation, slide 23
Think about that from the customer's perspective.
The infrastructure provider reduces its upfront funding requirements by collecting cash earlier.
The customer takes on additional liquidity and demand risk.
For a hyperscaler, this can involve billions of dollars. For a seed-stage startup, it might involve $400,000.
The contracts are not identical, but the financial mechanism is similar.
The Contract That Only Works If the Pitch Deck Comes True
Consider a hypothetical AI startup preparing for its Series A.
Current monthly revenue is $100,000. Current compute consumption costs $20,000 at standard usage-based prices.
The company's fundraising model forecasts rapid customer growth and a 15% monthly increase in compute demand.
Under that forecast, compute usage over the next 12 months will cost approximately $580,000 at today's standard prices.
The infrastructure provider offers a 25% discount in exchange for a $400,000 upfront prepayment.
The credits expire after 12 months.
At discounted prices, the prepayment covers approximately $533,333 of standard-price compute usage, or about 92% of the company's forecast demand. Any consumption beyond the prepaid balance is billed at standard rates.
Now the deal looks reasonable.
The startup expects enough growth to consume its credits. It anticipates substantial savings.
But the contract doesn't adjust if its forecast turns out to be wrong.
That's the trap.
The company has effectively turned its growth forecast into a financial commitment.
Three Futures for the Same $400,000 Decision
Not every AI startup buys compute this way.
Companies using pay-as-you-go APIs generally pay for actual consumption. Others purchase prepaid credits, reserve GPU capacity, or sign minimum-spend agreements.
Those arrangements have different financial consequences. A future minimum commitment is not the same as an immediate cash payment.
Our example assumes a $400,000 upfront payment for flexible compute credits that expire after 12 months. There are no monthly minimums, additional penalties, or refunds. Eligible usage is charged at 25% below standard prices.
We hold the contract terms constant and change only compute consumption.
Hypothetical 12-month comparison
| Metric | Growth: +15% monthly | Flat usage | Decline: -8% monthly |
|---|---|---|---|
| Pay-as-you-go compute cost | $580,033 | $240,000 | $158,083 |
| Prepaid credits consumed | $400,000 | $180,000 | $118,563 |
| Prepaid asset after month 6 | $268,694 | $310,000 | $326,192 |
| Credits forfeited at expiration | $0 | $220,000 | $281,437 |
| Total compute cash paid with commitment | $446,700 | $400,000 | $400,000 |
| Savings/(loss) vs. pay-as-you-go | +$133,333 | -$160,000 | -$241,917 |
All figures are rounded. Usage changes compound monthly from an initial $20,000 standard-price monthly compute requirement. The growth scenario includes approximately $46,700 of additional pay-as-you-go payments after credits are exhausted. The prepaid asset starts at $400,000, declines as services are consumed, and reaches zero after exhaustion or expiration. Write-off recognition may be required before expiration.
Scenario 1: Growth happens
Compute demand increases 15% every month.
The startup consumes the entire prepaid balance and pays approximately $46,700 for additional usage.
Total compute cash spending is approximately $446,700 instead of $580,033.
The contract saves approximately $133,333.
The economics work, provided the upfront payment doesn't jeopardize other financing needs.
Scenario 2: Growth stalls
Compute usage remains at $20,000 per month at standard prices.
Twelve months of pay-as-you-go compute would have cost $240,000.
At discounted rates, the startup consumes $180,000 of prepaid credits.
The remaining $220,000 expires.
The company paid $400,000 for services it could have purchased for $240,000.
Its 25% discount has turned into a $160,000 cash disadvantage.
Scenario 3: Compute demand declines
Usage falls 8% monthly because of fewer customer workloads, reduced model usage, or greater efficiency.
The standard-price annual compute requirement drops to approximately $158,000.
The company still pays $400,000.
Its cash disadvantage approaches $242,000.
And this is where the economics become counterintuitive: making the application more compute-efficient can increase the amount of prepaid capacity that expires unused.
One commitment. Three cash outcomes.
Annual cash cost · USD
Growth +15% monthly
$133,333 saved vs. pay-as-you-go
Flat usage
$160,000 lost vs. pay-as-you-go
Decline -8% monthly
$241,917 lost vs. pay-as-you-go
Why the 85% Margin Misleads the Founder
Consider the flat-demand scenario.
The startup generates $1.2 million in annual revenue and consumes $180,000 of discounted compute.
Before considering unused credits, its compute-only margin on recognized consumption is:
($1,200,000 - $180,000) / $1,200,000 = 85%
But the company paid $400,000 upfront.
Its effective cash-based compute margin is:
($1,200,000 - $400,000) / $1,200,000 = 66.7%
That is not a GAAP gross-margin calculation. It is a management measure showing the actual compute cash burden relative to revenue.
The accounting reconciliation matters.
Initially, the $400,000 payment is assumed to create a prepaid asset. As eligible compute is consumed, $180,000 is recognized as expense over the year.
The remaining $220,000 must be assessed for recoverability and written off when appropriate if it cannot be used or recovered.
The timing and income-statement classification depend on the agreement and applicable accounting requirements.
If the write-off is classified in cost of revenue, the full-year compute-only gross margin would also fall to approximately 66.7%. If classified elsewhere, reported gross margin could differ, even though the economic loss remains.
This is why founders need three separate views:
- Gross margin: Revenue minus recognized delivery costs.
- Contribution margin: Revenue minus variable or avoidable costs of serving customers, consistently defined.
- Operating cash flow: Actual operating cash receipts and payments, including relevant prepayments and their timing.
An improving gross margin does not automatically mean an improving cash position.
The 75% Break-Even Test
The company prepaid $400,000 for the equivalent of $533,333 in standard-price compute capacity.
To break even against pay-as-you-go, it must consume at least $400,000 of compute at unchanged standard prices.
That represents 75% of the nominal capacity purchased.
Break-even utilization = Prepayment / Standard-price value of purchased capacity
$400,000 / $533,333 = 75%
Below that threshold, pay-as-you-go is cheaper.
Above it, the commitment can save money.
But even 75% utilization is not necessarily a good investment.
The formula excludes the opportunity cost of tying up $400,000, changes in market pricing, and the possibility of delaying a critical financing milestone.
The founder should model those risks before signing.
75% break-even utilization
$400,000 prepayment / $533,333 standard-price capacity = 75%
The Runway Cost Nobody Put in the Pitch Deck
Now consider a separate flat-demand runway illustration.
The startup has $1.2 million in cash. It collects $100,000 each month and pays $160,000 in monthly operating expenses excluding compute.
Without the commitment, it pays another $20,000 monthly for compute.
Monthly net cash burn is $80,000.
Runway: 15 months.
With the prepayment, the company immediately spends $400,000, leaving $800,000.
For the first 12 months, monthly cash burn is $60,000 because compute is covered by prepaid credits.
At the end of month 12, only $80,000 remains.
When ordinary compute billing resumes, monthly burn returns to $80,000.
Runway: approximately 13 months.
The commitment shortened runway by two months.
That difference could matter enormously if the founder needs 14 months to reach the milestones required for the next round.
The startup with the lower reported compute price could be forced to raise money before achieving the metrics needed for its target valuation.
The runway cost: 2 months lost
What If Compute Prices Fall Before the Contract Expires?
Technology creates another risk.
On November 6, 2026, Bloomberg reported that Microsoft was winding down an Oracle compute-rental arrangement involving older-generation Nvidia GPUs after the capacity no longer matched Microsoft's evolving requirements.
The report cited people familiar with the arrangement. It was not an announcement of a legal contract dispute.
Oracle disputed the broader implication that older-generation GPUs lacked demand.
In its October 2026 investor presentation, Oracle reported a 20% premium on certain renewed or resold capacity involving four-year-old GPUs.
Sources: Bloomberg, November 6, 2026 | Oracle Investor Day, slide 20
Both observations can be valid.
Hardware that becomes less attractive for one workload may remain valuable for another.
For a startup, the danger is signing an inflexible contract while its model architecture, hardware requirements, or provider pricing continue to change.
Consider another hypothetical.
Suppose standard pay-as-you-go compute prices fall 30% halfway through the year, while usage remains flat.
Annual pay-as-you-go spending falls from $240,000 to $204,000.
Assume prepaid dollar credits are charged at 25% below the then-current standard price. Lower prices make each credit dollar cover more compute, but the credits still expire after 12 months.
The company consumes only $153,000 of its prepaid balance. The remaining $247,000 expires.
It has still paid $400,000.
Its cash disadvantage increases to $196,000.
Falling prices benefited the flexible customer immediately. The prepaid customer could not recover the cash already committed.
The Founder Decision: Commit the Baseline, Not the Forecast
Some compute commitments make excellent financial sense.
A startup with stable, proven workloads may achieve meaningful savings through reserved capacity. Guaranteed access may also be valuable when compute availability is essential to meeting customer obligations.
AWS Savings Plans, for example, offer one- or three-year commitments with different upfront payment options.
Google Cloud's committed-use discounts can create noncancellable monthly payment obligations, regardless of actual usage.
Sources: AWS Savings Plans | Google Cloud committed-use discounts
These are different arrangements from our hypothetical prepaid-credit contract. Their cash schedules and break-even calculations need to be modeled separately.
Before signing, a founder should answer five questions:
-
What usage is already proven? Separate recurring production workloads from speculative customer growth and experiments.
-
What is the downside loss? Calculate the cash cost if usage remains flat, falls 20%, or becomes more efficient.
-
When does cash actually leave? Distinguish upfront payments from monthly minimums, future commitments, and cancellation penalties.
-
What flexibility survives? Can credits move between services, models, GPU generations, regions, or workloads? Do they expire? Are unused amounts refundable?
-
Which financing milestone is exposed? Recalculate runway after the commitment, not just compute cost per unit.
For many pre-seed and seed-stage founders, the better decision will be to commit the predictable baseline and keep uncertain demand on usage-based pricing.
This is particularly relevant when startup-program cloud credits expire.
Programs such as AWS Activate and Google for Startups provide promotional credits, but those credits have eligibility and expiration rules. They are not the same as cash-funded compute prepayments.
Better Gross Margins Don't Always Mean a Better Business
The dangerous assumption is that a lower compute price automatically improves a startup's financial position.
It doesn't.
A discount saves money only when the startup uses enough eligible capacity, at the right time, without sacrificing cash it needs elsewhere.
The startup in our example looked more profitable after signing its compute contract.
But if demand stayed flat, it finished the year with $160,000 less cash and approximately two fewer months of runway.
The contract was priced for the company in its pitch deck, not the company that actually existed.
Before committing a material portion of your runway to cloud infrastructure, build a model that shows expected savings, downside losses, payment timing, and financing consequences.
If your startup is considering a significant compute commitment or needs to evaluate how it affects the next funding milestone, speak with Andrew Financial about the decision.
Andrew Financial is the separate fractional CFO and financial advisory services business associated with Andrew. Andrew's AI-native software does not automatically evaluate or optimize cloud infrastructure contracts.
Before You Commit Your Next $400,000, Model the Downside.
A lower compute price doesn't automatically create a better financial outcome. Andrew Financial helps startup founders evaluate material infrastructure commitments against cash flow, runway, and fundraising milestones.
Discuss your compute commitmentResearch note: Public-company disclosures and reporting reflect information available as of November 14, 2026. All startup financial scenarios are hypothetical. Compute-only margins exclude other cost-of-revenue items. Cash-based compute margin is not a GAAP measure. Actual accounting recognition, asset recoverability, and cash-flow presentation depend on contract terms and applicable accounting standards.
