Agent Workloads Are Repricing the Infrastructure Beneath Them

The workload changed first

The core argument from Jump Capital is that the unit of work has changed. Agentic workflows involve long-horizon, multi-step tasks that challenge current inference infrastructure, whereas existing engines were built to optimise latency and throughput for individual requests. Batching requests improves aggregate throughput but can increase latency for individual users, a trade-off that matters more when a single agent task chains many steps together. m:…8320

Volume data supports the shift in workload shape. Jump Capital reported that OpenRouter weekly token volume grew 250 times from January 2025 to September 2026, and that agent traffic accounted for 7 times more tokens than human-driven usage on OpenRouter. Enterprises are aggressively adopting open-weight models, according to Jump Capital, and inference providers such as Baseten, Together and Fireworks have scaled rapidly to support demand. The implication for capital allocation is that token growth is increasingly agent-driven, not user-prompt-driven. m:…8320

Jump Capital describes the inference market as a 130 billion dollar market today, on track to hit 350 billion dollars by 2027. That projection is Jump Capital's own estimate, not a reported market total, and it should be read as a statement of the scale of the opportunity as Jump Capital sees it. Even discounted, the direction is clear: a market that triples in two years on Jump Capital numbers would reward suppliers who can serve multi-step work, not only fast single answers. m:…8320

The funding gap inside enterprises

Enterprise demand is visible, but enterprise funding is not keeping pace. Intel says IDC research commissioned by Intel found 89 percent of information technology decision makers expect agentic artificial intelligence to be important to their responsibilities within the next 12 to 24 months, while only 56 percent of organisations say agentic infrastructure is well or adequately funded. Intel describes infrastructure funding as a significant bottleneck, and 45 percent of respondents cited modernising or consolidating data centre infrastructure as a top information technology priority. m:…6222

Deployment preferences add a second constraint. Intel reported that 55 percent of respondents take a balanced approach between private information technology and public cloud, while 38 percent primarily or exclusively prefer private information technology, and approximately 40 percent of buyers prefer standard servers for agentic solutions. A market split between private and public infrastructure complicates any single-vendor answer and spreads spending across hardware, software and managed services rather than concentrating it in one layer. m:…6222

Jump Capital argues that the technical requirements explain the funding tension. Agentic inference requires end-to-end co-design of graphics processing unit kernels, compilers and runtimes, persistent execution state instead of temporary caches, and workflow-aware scheduling. Current engines such as vLLM and SGLang use continuous batching, paged attention and prefix caching, techniques designed for request-level optimisation. Rebuilding for workflow-level optimisation is a capital project, not a configuration change, which is why the funding gap matters to suppliers and buyers alike. m:…8320

The data layer becomes the balance sheet item

Databricks illustrates where one supplier is placing its bet. Andreessen Horowitz describes Databricks as having surpassed a 7 billion dollar revenue run rate and growing at 80 percent year over year, with more database queries from agents than from humans and agents launching more than 17 million databases on Databricks every day. Those figures come from Andreessen Horowitz account of the company, and they frame the data platform as the system of record for agent work rather than a passive store. m:…1875

The commercial logic follows from grounding and governance. Andreessen Horowitz reported that grounding an agent in a company ontology raised answer accuracy from 50 percent to 84.5 percent in the example it cited, and that Databricks Lakebase product has crossed a 100 million dollar revenue run rate. Andreessen Horowitz argued that sustainable competitive advantage in an agentic firm requires owning the data and decision loops, a claim that, if accepted, makes data architecture a capital decision rather than an information technology detail. m:…1875

Venture funding is following the same thesis at an earlier stage. Jump Capital led a 20 million dollar seed round for Standard Kernel, a company working on inference infrastructure, consistent with a market in which the constraint has moved from models to the systems that run them. For professionals allocating capital, the question is no longer whether agent workloads grow, but which layer captures the value: the model, the inference engine, or the governed data platform underneath both. m:…8320