The quiet infrastructure story behind the AI cost debate
While the trillion-dollar IPOs grab headlines, one of the most consequential funding rounds of the week went to a 14-person company. Ollama raised a $65 million Series B led by Theory Ventures, following a $15 million Series A led by Benchmark's Peter Fenton, for $88 million raised in total. The traction numbers explain the investor enthusiasm: launched in 2023 to make open-weight models runnable on an ordinary PC in minutes, Ollama now counts 8.9 million monthly developers, presence in 85% of the Fortune 500, and 176,000 GitHub stars with nearly 17,000 forks.
Docker's playbook, replayed for AI
Founders Jeff Morgan and Michael Chiang know exactly what product ubiquity among developers looks like - they helped build Docker Desktop after Docker acquired their startup Kitematic. The analogy is almost literal: Docker abstracted away hardware configuration so cloud apps could run anywhere; Ollama does the same for open models, which in 2023 were built for researchers and painful for ordinary programmers to stand up. The monetisation layer follows the same freemium logic: the free desktop tool remains untouched, while a cloud service hosts larger models under subscription tiers from free to $100 per month - notably metered on GPU time rather than token limits, a pricing model worth watching as buyers grow weary of opaque token math.
Why now: the open-model inflection
Morgan pins the business inflection to January, when agentic assistants took off and large open models suddenly proved capable of real work like coding. That fuels the industry's sharpest cost argument: Fenton frames it as no longer either/or between open and closed models, but says any company with high inference expenses now has a vital, existential project to shift workloads toward open weights. Ollama is one of a whole crop of open-source projects turning into venture-backed companies - inference providers like Inferact (vLLM) and RadixArk (SGLang), assistant alternatives like NanoClaw, and small model builders like Arcee. There is friction, too: parts of the community accused Ollama of drifting toward commercialisation a year ago, criticism the founders answer by insisting the free local product is unchanged.
Where the money is for you
- The practical takeaway is a hybrid model strategy: route routine, high-volume workloads (classification, extraction, internal chat, first-pass drafting) to open models served locally or via commodity inference, and reserve frontier closed models for the tasks that genuinely need them. Companies making this split are the reason Ollama sits in 85% of the Fortune 500.
- Local execution is also a privacy and compliance play: data that never leaves the machine sidesteps a whole category of vendor-risk review, which is often what unblocks AI pilots in regulated teams.
- If you advise clients, an inference-cost audit is now a legitimate standalone engagement: map which workloads run on premium closed models today, benchmark open-weight equivalents, and quantify the delta - the existential framing from Ollama's own board suggests the savings are frequently material.
- One caution for procurement: a 14-person vendor is a thin organisation to bet critical infrastructure on, however impressive the penetration. The mitigation is the product's open-source core - your exit path is the code itself, which is precisely why this model of company keeps winning enterprise trust.