Will cheaper AI inference weaken NVIDIA or expand its market?

NVIDIA’s operating income for the quarter ended July 26, 2026 was $63.734 billion, up from $53.536 billion in the previous quarter. Meanwhile, Amazon says its own Trainium chips already run most Bedrock inference. NVIDIA’s earnings growth and customers’ use of alternative chips are happening together.

In that quarter, NVIDIA’s company revenue rose 17.9% sequentially and operating income 19.0%. Gross margin stayed near 75%, close to the previous quarter. Revenue, gross profit and operating income all increased across five quarters. The company statements do not yet show cheaper computation squeezing operating earnings. They cover more than inference and do not measure GPU units.

Lower computation costs can make previously uneconomic AI features viable or fund deeper reasoning within an existing budget. How many orders NVIDIA gains depends on how much computation those uses need and which systems customers choose.

Within a consistent billing boundary, a customer’s total spending is average paid price times paid quantity. If quantity growth more than offsets the price reduction, the bill can still rise. Longer contexts, additional answer candidates and verification can also consume more computation per job. A cheaper unit does not necessarily mean a smaller service budget.

API prices are not equipment requirements. Smaller models, more efficient computation, cheaper hardware, better utilization and provider concessions can all reduce the bill. Some change the resources needed for the same work; a concession may not. Compare the full resources required to meet the same quality standard, including failures and retries.

The calculator below illustrates resources only. Assume each task uses 40% less and task count grows 60%: the resource index becomes 96, still 4% below the starting level. Offsetting that saving requires about 66.7% more tasks. These are adjustable assumptions, not an API discount or measured demand growth.

NVIDIA’s revenue also depends on the customer’s hardware choice. Additional computation may run on GPUs, TPUs or Trainium, with different product prices and margins. More calls, more computation and higher NVIDIA earnings are not interchangeable results.

In its July 22, 2026 Q2 CEO remarks, Google reported model API processing at about 22 billion tokens per minute, compared with roughly 16 billion in the previous disclosure. The rounded figures imply an increase of about 37.5%, and Google still described supply constraints. These are rates at two disclosure points, not cumulative quarterly usage.

In a separate service, Google said AI Mode response costs were the lowest since launch while overall Search gained incremental queries. API and Search measures cannot form a single price–demand curve. They nevertheless show an operator reporting improved efficiency alongside expanding use, rather than an inevitable contraction as costs fall.

Commercial use extends beyond trials. Google also reported nearly 500 Cloud customers each processing more than one trillion paid tokens over the preceding twelve months. Token counts vary with model and workload mix and cannot be converted into GPU purchases. They remain closer to actual demand than product launches, downloads or financing announcements.

Deeper reasoning can consume some of the budget that efficiency releases. OpenAI’s 2024 o1 research showed improved performance on some reasoning tasks with more computation when generating answers. Customers can spend less on the same work or use the same money to seek better answers. Existing paid usage supports expansion more convincingly than price cuts alone support a smaller compute market.

Amazon says Trainium already runs most Bedrock inference. It also says Claude models use more than one million Trainium2 chips across training and operation. The first statement concerns Amazon’s own platform; the second spans two kinds of computation. Neither measures global inference share, but both describe deployment rather than an unfulfilled roadmap.

Claude challenges the split between dedicated chips for simple models and GPUs for the smartest ones. Whether customers can make dedicated chips work depends on request volume, sustained workloads, and their ability to optimize models, compilers and chips together. At sufficient scale, long-term engineering work can be spread across more services. AWS has those organizational capabilities and an operating deployment.

For NVIDIA, the specific competition is that such buyers can route more new work to their own chips. Reusable GPU tools still matter to customers with changing models, mixed workloads and urgent deployment needs. Less low-level adaptation may be worth more than a cheaper individual chip. Engineering and service costs divide these settings, not model sophistication.

Large cloud providers need not choose only one architecture. Google offers both its own and NVIDIA accelerators and supports software portability between GPUs and TPUs. A growing market can support several routes. As work becomes mature and repetitive, customers have more scope to compare platform costs. The pressure on NVIDIA is competition for new orders and pricing, not AI suddenly losing its uses.

NVIDIA’s development documentation describes two sets of tools: its Triton backend batches requests, reuses cached results and supports execution across devices; Dynamo plans capacity around request throughput and waiting-time targets. These tools aim to turn GPU computing capacity into a usable service.

KV caching saves intermediate results for reuse during later generation. Waiting-time targets cover how soon output starts and the intervals between subsequent outputs. The documentation describes functions, not measured customer savings. For customers repeatedly deploying new models, the tools’ value depends on how much adaptation they avoid and whether they meet service requirements.

Meta supplies a specific counterexample. It reported DDA communication optimization bringing MI300X to overall H100 performance parity in its particular test configuration. That does not settle every workload comparison. It shows an engineering-capable buyer improving an alternative instead of paying NVIDIA for all the integration. AWS’s in-house deployment is another route that depends on organizational capabilities.

NVIDIA must keep making its systems worth retaining when customers change models or add services. Faster deployment, less operating work and reliable completion can preserve orders. Customers will compare those benefits with system prices. A software ecosystem is a competitive asset, not a guarantee that profit remains safe.

NVIDIA’s operating income rose 19.0% sequentially in the quarter, but net income increased only 2.3%. Looking only at the bottom line can mistake nonoperating swings for a slower chip business. The difference largely comes from changes outside operating earnings.

Operating income increased $10.198 billion from the previous quarter; other income net fell $8.594 billion and tax expense rose $0.237 billion. Net income consequently increased $1.367 billion. The bridge below shows how lower nonoperating income held back net-income growth without establishing a stagnant chip business.

Cash warrants separate attention. Q2 operating cash was $24.077 billion, down 52.2% sequentially but up 56.7% year over year. Receivables grew 54.9% sequentially, faster than revenue’s 17.9%. Recorded sales and profit increased without cash keeping pace. Collections, contract terms, company payment timing and working capital all matter to that difference.

Profit and cash moved differently this quarter, but the cash decline does not establish a shift to alternative chips. Turning profit into cash depends on customer collections, contract terms and NVIDIA’s own payments. One quarter’s cash decline does not establish shrinking inference demand.

Through the latest disclosed quarter, NVIDIA revenue and operating income were still growing, as was Google’s model API usage. These records do not show a simple outcome in which cheaper computation shrinks the business, nor do they establish lower costs as the cause of growth. For NVIDIA, the more useful question is whether that additional work will still lead customers to buy its systems.

Amazon’s operating Trainium deployment offers a concrete countercase. If large customers route more new inference work to in-house chips while NVIDIA’s gross profit and operating income remain under pressure, growing paid usage would not rescue the case for its system advantage. New orders continuing to choose NVIDIA alongside sustained operating-profit growth would strengthen that case. Whether customers pay within their contract terms requires a separate check.