Summary

  • Snowflake committed to at least US$6.0 billion of AWS infrastructure spend over five years ending 31 March 2031. The arithmetic average is US$1.2 billion a year, but the actual annual minimums are undisclosed amounts ranging from US$900 million to US$1.25 billion.
  • Missing either an annual floor or the cumulative minimum requires Snowflake to pay the difference. Those payments can be applied to qualifying cloud use during the contract term, so a miss can shift the timing of economic use rather than become an immediate total loss.
  • The amendment removed the remaining obligation under the predecessor form of the agreement without stating its amount. Snowflake's earlier US$2.681885 billion aggregate commitment table covered several cloud and enterprise contracts, so subtracting it from US$6 billion would create a false incremental number.
  • Q1 cloud infrastructure expense was about 73% of US$386.874 million of product cost, or roughly US$282.4 million across all providers. That makes the new AWS floors economically material, but it does not disclose AWS-only utilization or prove either underuse or margin pressure.

The easiest way to misunderstand Snowflake's new AWS agreement is to divide US$6 billion by five, call the answer a run rate and stop. The division is correct: US$1.2 billion. The contract is not that simple.

Snowflake's April 2026 filing places two controls over the same five-year term. There is a cumulative minimum of US$6 billion through 31 March 2031. There is also a minimum for each contract year, with annual amounts somewhere between US$900 million and US$1.25 billion. Snowflake does not publish the sequence.

That missing sequence matters. A back-loaded workload plan may still be on course to clear US$6 billion overall and yet miss an early annual floor. A front-loaded plan may clear the early tests while leaving less room for a later change in architecture. The headline measures total commercial commitment; the annual floors measure timing discipline.

This is the economic receipt investors need to follow. Snowflake sells customers a consumption platform whose revenue follows use. It has now accepted a supplier contract that forces part of its own cost base to keep time even when workload demand does not.

One contract contains two minimums

The five-year arithmetic permits many possible annual schedules. Five years at US$1.2 billion would reach US$6 billion. So would a sequence that begins near US$900 million and later rises toward US$1.25 billion, provided the exact amounts sum to the cumulative requirement. The public filing does not say which pattern applies.

It would therefore be wrong to call US$1.2 billion the first-year floor, next year's floor or even a contractual floor at all. It is only the cumulative amount divided by the term. The disclosed yearly range is the harder information because it shows that Snowflake cannot freely move all unused spend to the final year.

The agreement also creates two ways to owe a difference. Snowflake must pay if it misses a contract-year minimum, and it must pay if it misses the cumulative minimum. The filing does not disclose whether an earlier payment reduces a later calculation dollar for dollar, how credits are consumed across services, or which price schedule applies when the paid balance is finally used.

What it does disclose is important: a shortfall payment can be applied to qualifying cloud-infrastructure spend during the contract term. That means payment and economic consumption can separate. Cash may leave before the corresponding compute, AI or other eligible service is used.

Calling every shortfall dollar “waste” would be premature. Calling it harmless would be equally weak. A credit that can still be used has option value, but that value depends on eligibility, expiry, workload location, negotiated price and whether future demand arrives before the contract ends.

The old balance disappears without becoming measurable

The April amendment also says Snowflake no longer has to fulfil the remaining non-cancelable purchase commitment under the predecessor agreement. It does not give that old balance.

Three months earlier, Snowflake's annual report showed US$2.681885 billion of future minimum payments under non-cancelable purchase commitments longer than one year. The table included third-party cloud infrastructure agreements and enterprise subscriptions. It was an aggregate measured on 31 January, not a predecessor balance measured on the April amendment date.

The annual report described two cloud arrangements more closely. One had US$530.5 million remaining under a US$1.0 billion commitment running from June 2023 to May 2028. Another had US$518.0 million remaining under a US$530.0 million commitment running from November 2025 to October 2030. Neither provider was named, and both contracts were described as having no annual minimum.

Nothing in the April filing says that the AWS amendment replaced either disclosed balance. Nor does it say that every other item in the January table was amended. US$6.0 billion minus US$2.681885 billion, therefore, is not US$3.318115 billion of new obligation. It subtracts a multi-contract, earlier-date aggregate from one amended, later-date agreement.

The uncertainty is not a defect to be filled with arithmetic. It is a monitoring item. A later commitment table should show the new yearly payment schedule alongside whatever obligations remain outside AWS. Until then, the defensible statement is that the contract perimeter grew or changed substantially, while the exact incremental bridge is undisclosed.

The supplier clock is different from the customer clock

Snowflake recognizes product revenue when customers consume compute, storage and data-transfer resources. It does not normally recognize that revenue ratably over the customer's contract term. Customers can consume beyond contracted capacity and may roll unused capacity into a later period, generally when they purchase more capacity on renewal.

This model makes customer activity variable by design. In Q1 FY2027, product revenue was US$1.334329 billion, 33.9% above the prior-year quarter. About 97% of total revenue came from existing customers under capacity arrangements. Net revenue retention was 126%.

The supplier agreement points the other way. AWS receives a cumulative minimum and an annual timing constraint. Snowflake receives the ability to apply qualifying spend and, according to its public announcement, deeper product integration, migration support and joint go-to-market activity. Customer use may accelerate, slow or move across clouds; the AWS floor remains contractual.

That mismatch is not automatically bad. A growing consumption business can use a purchase commitment to obtain volume discounts, secure commercial attention and support wider regional deployment. Snowflake says higher volume discounts helped keep Q1 product gross margin flat at 71% even as cloud infrastructure expense increased.

But the mismatch transfers forecasting risk to Snowflake. It must estimate not just total platform demand, but the portion eligible for one provider in each contract year. Multi-cloud customer choice, data-residency requirements, GPU availability, product architecture and price-performance can change that share.

Current expense shows scale, not coverage

Snowflake reported US$386.874 million of product cost in Q1 FY2027. Third-party cloud infrastructure expense represented approximately 73%, which implies about US$282.4 million for the quarter. Multiplying by four gives an annualized ruler near US$1.130 billion.

The five-year commitment's US$1.2 billion arithmetic average is about 6.2% above that ruler. The comparison makes the commitment's scale visible. It cannot show whether Snowflake is ahead of or behind contract.

The expense estimate covers AWS, Azure and Google Cloud, different regions, customer workloads, GPUs, AI inference and platform deployment. The commitment belongs to AWS. The actual first-year floor may be US$900 million, US$1.25 billion or an amount between them. Q1 seasonality and growth also make a four-times annualization a poor forecast.

The prior year offers a second ruler. FY2026 product cost was US$1.260324 billion and cloud infrastructure expense was approximately 71%, or about US$894.8 million by mechanical calculation. The US$1.2 billion average is 34.1% higher, but once again one figure is recognized all-provider historical expense and the other is a forward average for one provider.

The useful question is not whether one ratio looks high. It is whether eligible AWS use, realized unit prices and product gross-profit dollars progress together. In Q1, cloud infrastructure expense rose US$86 million year over year, mainly because of customer consumption, while product gross margin stayed at 71%. New capabilities that had not reached scale offset some discount benefits.

That coexistence is healthy counterevidence. More cloud cost accompanied much more product revenue and stable margin. It still does not guarantee that every future annual floor will be absorbed economically.

RPO cannot be used to pay a provider

Snowflake ended Q1 with approximately US$9.21 billion of remaining performance obligations, up 38%. It is tempting to place that customer-side number beside US$6 billion and call the difference coverage.

The two ledgers do different jobs. RPO is contracted future revenue that has not yet been recognized, including deferred revenue and non-cancelable amounts not yet billed. It is gross customer consideration, not cash already collected, gross profit or AWS-eligible workload. Snowflake also says the timing of consumption by any particular customer is hard to predict.

The US$6 billion figure is a supplier-side spending minimum. It may support the compute and AI needed to produce revenue, but it is not an offset inside RPO. A customer can hold a Snowflake commitment and consume on Azure. An AWS workload can have a different margin by product and region. Marketplace sales can change procurement without matching Snowflake's expense timing.

The public collaboration announcement adds another figure: more than US$7 billion of lifetime Snowflake sales through AWS Marketplace, including more than US$2 billion in calendar 2025. Those are valuable distribution receipts. They are not supplier-spend receipts. Treating marketplace sales as AWS-cost utilization would merge customer procurement with Snowflake's input bill.

Liquidity is a counterweight, not an answer

Snowflake had approximately US$4.4 billion of cash, cash equivalents and investments at 30 April. The five-year commitment is 36.4% larger than that point-in-time liquidity. This ratio sounds alarming only if the entire commitment is treated as immediately payable.

It is not. The contract runs to March 2031, and ordinary eligible usage should satisfy part of the minimums. Q1 operating cash flow was US$243.223 million. Non-GAAP free cash flow was US$232.772 million, and adjusted free cash flow was US$265.514 million. Those amounts argue against an immediate funding emergency.

They do not remove the economic choice. Cash used to cover a shortfall cannot be used for acquisitions, repurchases or another cloud provider at the same time. Snowflake spent US$300 million on repurchases during Q1 and completed the Observe acquisition with substantial cash and shares. Capital allocation must now coexist with annual supplier floors.

The strongest reading is neither distress nor triumph. Snowflake bought a large, timed input commitment while its consumption business was growing 34% and holding product margin. The agreement can reinforce scale economics if workloads arrive in the right provider, product and year. It can narrow choice if they do not.

Sources