The Short Answer
AI infrastructure now pays for itself — but the return lands on the balance sheet of whoever owns the stack. In the quarter ending June 30, 2026, AWS grew 37% to $42.2 billion, Google Cloud grew 82% to $24.8 billion, Azure crossed $100 billion in annual revenue, and Microsoft 365 Copilot passed 30 million paid seats. Every one of those seats is rent.
For a buyer, the strategic question is no longer whether to spend on AI. It is whether that spend converts into an owned asset or into someone else's recurring revenue.
Per-seat licensing guarantees the latter: the bill multiplies by headcount regardless of use, and none of it accrues to you. Usage-based and self-hosted deployment invert that — you pay for work performed, on infrastructure and source code you control, running any model you choose.
What did Q2 2026 earnings actually prove about AI infrastructure?
They proved the capex cycle has turned into revenue, at a scale that removes any remaining ambiguity. The four largest providers all reported acceleration, not maturation, in the quarter ending June 30, 2026.
Amazon Web Services reported $42.2 billion in quarterly revenue, up 37% year over year — its fastest growth in 18 quarters. CEO Andy Jassy disclosed that Amazon's AI business and its custom silicon business have each passed $25 billion in annualized run rate.
AWS is not only hosting the workloads; it is selling the chips underneath them.
Google Cloud posted $24.8 billion, up 82% year over year, with CEO Sundar Pichai noting that Gemini Enterprise has reached nearly 90% of the Fortune 100.
Microsoft reported $90 billion in total quarterly revenue, with Azure and other cloud services growing 43% and Azure crossing $100 billion annualized for the first time.
Meta generated $60.8 billion, up 28%, and made the most revealing disclosure of the four: CEO Mark Zuckerberg confirmed the company is evaluating leasing its excess AI compute to third parties at prices carrying what he called a significant premium over acquisition cost.
Meta has accumulated so much AI infrastructure that reselling the surplus is now a candidate business line.
Who captures the return on enterprise AI spending?
The infrastructure owner does, and the earnings make the mechanism explicit rather than implied. Each of these companies monetizes AI because it holds the full stack — silicon, training pipeline, inference capacity, and the distribution channel that reaches the buyer.
That is not a criticism of the vendors; it is a description of where value settles. AWS grew its custom silicon into a $25 billion business by making AI workloads cheaper to run on its own hardware than anywhere else.
Google's 82% cloud growth arrives alongside a Fortune 100 penetration figure near 90%, which is a dependency statistic as much as an adoption one. Microsoft's 30 million paid Copilot seats are 30 million recurring reasons not to leave.
When an enterprise buys tokens, fine-tunes on a managed platform, or provisions seats, the resulting capability is real — and the durable asset created belongs to the vendor. The buyer's spend shows up as operating expense that resets to zero every renewal cycle.
The vendor's shows up as infrastructure that compounds.
How much does per-seat AI actually cost a 10,000-employee enterprise?
Far more than the same workload costs when priced by usage, and the gap is arithmetic rather than opinion.
Microsoft 365 Copilot at roughly $30 per user per month across 30 million seats is approximately $10.8 billion in annualized revenue — which means the per-seat model is now the single largest transfer of enterprise AI budget in the market.
Scale that down to one organization and hold the workload constant. Assume 10,000 employees, of whom 2,000 use AI actively at about 20 requests per working day — roughly 10 million requests a year, at approximately 3,000 input and 700 output tokens each.
| Pricing shape | Unit price | Annual cost | Scales with |
|---|---|---|---|
| ChatGPT Enterprise (per seat) | ~$60/user/mo | ~$7,200,000 | Headcount |
| Glean (per seat) | ~$40/user/mo | ~$4,800,000 | Headcount |
| Microsoft 365 Copilot (per seat) | ~$30/user/mo | ~$3,600,000 | Headcount |
| Token-priced, frontier model | ~$3 / $15 per MTok | ~$195,000 | Requests |
| Token-priced, mid-tier model | ~$1 / $5 per MTok | ~$65,000 | Requests |
| Self-hosted, open weights on owned GPUs | 1 × 8-GPU node | ~$120,000–180,000 | Nothing — flat |
Every figure above is the same workload. The only variable is the shape of the contract.
Per-seat multiplies a fixed monthly fee by every employee who has access, whether they run 20 requests a day or none; the bottom three rows bill for work actually performed, or for hardware that costs the same at 2,000 users as at 20,000.
That is why per-seat is the wrong shape rather than a more expensive option. A 10,000-person enterprise can double its AI usage under token pricing and still pay a fraction of the seat bill — and if it grows to 15,000 employees without changing usage, the seat bill grows 50% for nothing.
We ran the same arithmetic for higher education in the true math of per-seat AI on campus.
What does it mean to own your AI infrastructure instead of renting it?
It does not mean building data centers or training foundation models. Owning the stack means holding the four things a subscription never conveys, each of which changes the economics above.
You own the source code. When the platform is code you hold under license rather than a seat you rent, every improvement your team makes is a capitalizable asset instead of a feature request in someone's backlog. See ownership versus rental as a cost model.
You choose the models. A model-agnostic architecture routes each workload to whatever performs best for it — a frontier API for hard reasoning, an open-weight model for high-volume or cost-sensitive work, a locally hosted model where data cannot leave.
Switching is a configuration change, not a migration. Given that open-weight models now match frontier-class performance, that optionality is worth real money every quarter.
You keep the data. Inference inside your own perimeter means institutional data never trains an outside model and never becomes leverage in a renewal negotiation.
You pay for usage, not headcount. Credit-based consumption ties cost to delivered value. Adding 3,000 employees to the directory does not add a dollar until those people actually run workloads.
Why does concentration in four vendors make this urgent for regulated buyers?
Because the same numbers that read as growth to an investor read as concentration risk to a regulator.
AWS at 37% growth, Google Cloud at 82%, and Azure past $100 billion annualized describe a market where the critical AI capability of most institutions runs on infrastructure controlled by four American corporations whose pricing, terms, and model availability can change on a quarterly cadence.
For an agency subject to data residency rules, a bank under examination, a health system with patient records, or a defense contractor under export controls, that dependency is a live compliance exposure rather than a hypothetical one.
Meta's disclosure that it may lease surplus compute at a premium is a useful illustration: capacity, and its price, is set by the owner's strategy, not the tenant's roadmap.
The structural answer is to hold the parts that must not change hands — the code, the data, and the deployment target — while treating the model as a swappable component. ibl.ai is built for exactly that posture, deploying in your cloud, your VPC, on-premise, or fully air-gapped, with the full stack under your control.
It also matters who you are buying that posture from. ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned partner with no outside owner whose incentives can reset your terms at the next funding round or acquisition.
What should enterprise leaders do before the next earnings cycle?
Treat AI spend as a capital allocation decision, and test it against the numbers this quarter produced. Four checks are enough to know which side of the ledger you are on.
Price your actual workload, not your headcount. Count requests and tokens for a representative month, then compare that figure against your per-seat invoice. If the gap resembles the table above — millions against low six figures — the contract shape is the problem, not the technology.
Audit your model dependency. If an application cannot switch models without an engineering project, you cannot capture the savings when the next open-weight release lands, and open-weight releases are now arriving monthly.
Establish where inference runs. For every workload touching regulated data, know whether the reasoning happens inside your perimeter. If it does not, that is a compliance item, not an architecture preference.
Decide what you want to own in three years. The organizations that matter in this market will not be the ones with the largest AI subscription line. They will be the ones that converted the same budget into owned code, chosen models, retained data, and costs that track usage.
Q2 2026 proved AI infrastructure pays off. The remaining question is whose balance sheet it strengthens.
ibl.ai is an Agentic AI Operating System that organizations deploy on their own infrastructure with full source code and data ownership — model-agnostic, usage-based, and deployable anywhere from managed cloud to fully air-gapped. Family-owned and operated from New York, NY. Learn more about enterprise deployment.