Test-time compute Market Outlook 2026–2033: Growth Drivers, Key Companies, Strategic Developments & Adoption Forecast — A Meticulous Next™ Foresight Brief

This Meticulous Next™ brief examines test-time compute, the extra computing a reasoning model spends while answering a question by generating long chains of intermediate tokens before it replies. It tests how far reasoning turns inference from a small cost line into the largest cost of AI by 2033, and what that means for cloud providers, chipmakers, model developers, software vendors and the companies that pay the bill.

One answer can cost many times more. OpenAI priced o1-preview at 60 dollars per million output tokens in September 2024 and o1-pro at 600 dollars in March 2025, and a December 2024 report put the difference between high and low compute settings of o3 at about 170 times, with the full ARC-AGI benchmark costing more than 10,000 dollars to run at the high setting. The conclusion is that the cost of a query is now a choice made at run time, which makes compute planning a product decision and not only an infrastructure decision.

Price cuts have not removed the pressure. DeepSeek released R1 in January 2025 at 0.55 dollars per million input tokens and 2.19 dollars per million output tokens, and per-token prices for models of fixed quality have kept falling, but reasoning models generate far more tokens for each query. The conclusion is that a lower price per token is offset by a larger number of tokens, so total inference demand rises even as unit cost falls.

Spending on capacity is very large. Capital spending by Amazon, Microsoft, Google and Meta rose from about 225 billion dollars in 2024 to about 352 billion dollars in 2025, and guidance for 2026 is reported at about 660 billion to 690 billion dollars, although these figures cover all cloud infrastructure and not AI alone. The conclusion is that the largest buyers are building capacity ahead of demand, which makes the cost of serving reasoning workloads a central question for returns on that investment.

The software layer is where cost can be cut. Selective use of reasoning, routing of easy queries to small models, caching and faster decoding reduce the tokens and the time spent on each answer, and Mem0 and similar vendors report token savings of up to 90% from selective retrieval. The conclusion is that buyers will pay for software that lowers the cost per answer, since each saved token is saved margin.

The brief treats test-time compute as a high-impact signal with a seven year horizon to 2033. It identifies which vendors capture value as reasoning and agent workloads become the majority of inference. It is a focused 112-page decision brief that presents an indicative trajectory rather than a segmented market model.

Reasoning moves compute from training to the moment of use. A reasoning model spends extra tokens thinking through a problem before answering, and giving it more tokens at run time improves accuracy on hard tasks. The conclusion is that quality can be bought per query, so the cost of AI follows the difficulty of each question and no longer follows the size of the model alone.

The gains are large on hard problems. DeepSeek reported that reinforcement learning raised accuracy on the AIME mathematics test from 15.6% to 71% and to 86.7% with majority voting, and these are company results. The conclusion is that extra thinking pays off most on maths, code and planning, which are the tasks that agents depend on and that enterprises will pay more to get right.

Memory and speed set the cost. Long chains of tokens fill the key value cache, so serving cost depends on memory capacity, memory bandwidth and the efficiency of decoding, and research such as ThreadWeaver reports average speedups of 1.53 times in token latency. The conclusion is that memory suppliers, network vendors and serving software capture a larger share of the cost than in earlier generations of models.

Routing decides the economics. A provider that sends every question to a high-effort reasoning model spends far more than one that sends easy queries to small models and reserves deep reasoning for hard ones. The conclusion is that routing, budgets and caching become core products, and customers will compare providers on cost per correct answer rather than cost per token.

The limits are rising total cost despite falling unit prices, shortages of memory and power and the difficulty of measuring when extra reasoning adds value. The brief tests each limit against the evidence to 2033.

The test-time compute market, covering inference optimisation software, serving platforms and managed services that run reasoning models and agents with long chains of generated tokens, and excluding GPU and accelerator hardware, rented compute capacity, model access fees charged per token and model training, is valued at USD 3.0–4.5 billion in 2026 and is projected to reach USD 19.86–29.79 billion by 2033, at a CAGR of 31%.

The 2026 base rests on licences, subscriptions and managed services for inference optimisation and serving sold to cloud providers, model developers and enterprises. Growth then follows the number of tokens generated by reasoning and agent workloads, the share of workloads that use cost-aware routing and the share of enterprises that buy serving software rather than build it.

The forecast is sensitive to the pace at which reasoning becomes standard in agents, the supply of memory and power and the degree to which models become more efficient in their thinking.

Scenarios.

The base case assumes steady growth in reasoning and agent workloads with continued gains in efficiency, bringing the inflection to about 2028 and a 2033 value near the middle of the range. An accelerated case assumes fast adoption of long-running agents that think for extended periods, moving the inflection to 2027 and lifting the 2033 value to the top of the range. A delayed case assumes models that reason with far fewer tokens and slower enterprise adoption, moving the inflection to 2030 and holding growth at the bottom of the range.

Growth drivers

• Spread of reasoning models in coding, analysis and agents.
• Large increase in tokens generated for each query.
• Capital spending by cloud providers on inference capacity.
• Enterprise pressure to control cost per answer.

Enablers

• Inference chips and memory with higher bandwidth.
• Serving software for caching and faster decoding.
• Routing between small and large models.
• Open weight reasoning models.

Restraints and barriers

• Shortages of memory, power and data centre space.
• Rising total cost despite falling unit prices.
• Difficulty of measuring the value of extra reasoning.
• Models that reason with fewer tokens.

The direction of test-time compute is set by the spread of reasoning models, the growth in tokens generated per query, capital spending on inference capacity, the supply of memory and power and the readiness of buyers to manage cost per answer.

The key participants in test-time compute and the infrastructure for reasoning models include NVIDIA Corporation, Advanced Micro Devices, Inc., Broadcom Inc., Alphabet Inc., Microsoft Corporation, Amazon Web Services, Inc., Meta Platforms, Inc., OpenAI, Inc., Anthropic PBC, Oracle Corporation, CoreWeave, Inc., Cerebras Systems Inc., Micron Technology, Inc., SK hynix Inc., Samsung Electronics Co., Ltd., Taiwan Semiconductor Manufacturing Company Limited, Dell Technologies Inc., Super Micro Computer, Inc., Databricks, Inc., Hugging Face, Inc., Intel Corporation, Qualcomm Incorporated, Arm Holdings plc, Marvell Technology, Inc., Cisco Systems, Inc., Arista Networks, Inc., Equinix, Inc., Digital Realty Trust, Inc., Nebius Group N.V., Lambda, Inc., Alibaba Group Holding Limited, DeepSeek, Together AI, Baseten, Fireworks AI, Groq, Mem0, Epoch AI, ARC Prize Foundation and the Linux Foundation.

The brief profiles representative participants in each archetype and assesses which are positioned to lead as reasoning and agent workloads become the largest part of inference.

NVIDIA Corporation, Advanced Micro Devices, Inc. and Broadcom Inc. lead among suppliers of inference chips and custom accelerators, Micron Technology, Inc., SK hynix Inc. and Samsung Electronics Co., Ltd. lead among memory suppliers whose bandwidth sets serving cost, and Microsoft Corporation, Amazon Web Services, Inc. and Alphabet Inc. lead among clouds that serve reasoning at scale. Competitive intensity is rising as chipmakers, clouds, neoclouds and software vendors all compete to lower the cost per answer.

Where value migrates.

In 2026 value sits with chip, memory and cloud suppliers that sell capacity for reasoning workloads, while software for cost control is still young. By 2028 it moves toward serving software, routing and caching products and toward memory suppliers, because buyers compare providers on cost per correct answer and memory bandwidth limits serving speed. By 2033 it settles with the clouds and model developers that run inference at scale with the lowest cost per answer, where scale, memory supply and software efficiency carry the margin. Margins hold for suppliers of memory and accelerators with supply secured and for clouds with scale, and narrow for providers of undifferentiated capacity that compete only on price.

Reasoning model price points, rising tokens per query, record capital spending by cloud providers and demand for cost control support growth, while shortages of memory and power, rising total cost despite lower unit prices and models that reason with fewer tokens hold back the pace.

A breakthrough in efficiency, a shortage of memory or power or a slowdown in enterprise agent adoption could change the pace. The brief therefore tracks four indicators: tokens generated per query, the share of workloads that use reasoning, the cost per correct answer and capital spending on inference capacity.

Reasoning everywhere first, then cost per answer as the main metric, then inference as the largest AI cost is the expected order of change.