Bloomberg Reports Increasing Number of Startups Building Custom Systems with Open Source Weight Models
Bloomberg reports that an increasing number of startups are building custom systems using open source weight models to cut costs and reduce reliance on OpenAI and Anthropic.
Legal tech company Harvey, valued at $15.6 billion, previously focused on specialized legal tasks using closed-source models like GPT-4. After a March update to its agents, customer usage surged, and insiders reported that gross margins fell from about 50% at the beginning of the year to approximately -50% by June. The company subsequently joined the ranks of those using open source weights, including models from Chinese labs, to create its own systems. In August, it released its proprietary model based on the Moonshot Kimi K3 training, with gross margins turning positive as usage adjusted. Harvey declined to comment on specific financials. The company claims token usage increased about 20 times that year, with two leading labs charging enterprises based on usage beyond subscriptions.
Similar actions have spread. Abridge is developing clinical models based on Nvidia's open source weights; Decagon claims about 80% of customer queries are handled by its proprietary models; Ramp raised $750 million in June, with the co-CEO stating that self-training made no sense a year ago but now does. Sequoia and General Catalyst are funding this chain. Anthropic reminded investors in an internal meeting that Harvey's most challenging tasks still require its flagship model.
Closed-source pricing based on tokens directly translates agent usage into cost curves; open source weights allow fine-tuning with proprietary data and deployment in proprietary systems, converting variable bills into fixed investments in computing power and engineering. Chinese open source weights have become an option due to low prices and availability for download, while in the U.S., companies like Arcee, Reflection, and Poolside are providing domestic open source alternatives.
Buyers are application-layer companies whose usage bills have eroded gross margins and must regain control over inference; sellers are labs expecting revenue solely from closed-source APIs. The event-driven aspect is that companies at the level of Harvey are reporting negative margins and model shifts in the same article. Beneficiaries include open source weights, inference clouds, and fine-tuning toolchains; those under pressure are cutting-edge closed-source interfaces that charge per usage.
The market trades on gross margins, not on who ranks first in models.
Source: Public Information
ABAB AI Insight
Application-layer companies first rent the strongest closed-source models to create products, then realize they have no pricing power after usage surges. Harvey transformed the lawyer workflow into an agent, and a March update caused the token curve to spike vertically, with negative gross margins indicating that rental costs are rising faster than customer prices. Switching to Kimi K3 for post-training is not about academic alignment; it removes the most expensive calls from the menu, leaving only the most challenging tasks for the closed-source flagship. This logic parallels Microsoft's shift to proprietary models for certain Office functions: diversify first, then internalize high-frequency, low-cost tasks.
The capital path is shifting from "subscription plus usage" to "buy weights, rent clusters, and maintain fine-tuning teams." Sequoia and General Catalyst are not funding just another shell chatbot but companies that can develop vertical models on open source foundations. Ramp states that self-training was not cost-effective a year ago but is now, with the dividing line being when usage surpasses the fixed costs of training and inference. Chinese open source weights serve as a price anchor, while U.S. open source labs sell "downloadable alternatives without geopolitical labels."
Comparative examples include companies that migrated core loads from AWS, manufacturers that switched from Salesforce to self-built CRMs, and smartphone manufacturers moving from Qualcomm reference designs to in-house developed basebands. The industry phase is shifting from "which model is smartest" to "which tasks are most expensive and can routing be disassembled." Closed-source labs still control the most challenging inferences but have lost monopoly over the high-volume, low-difficulty calls.
The structural judgment indicates a transfer of pricing power. Agents have transformed model usage from occasional Q&A to continuous work hours, with pricing units shifting from seats to tokens. The mechanism is: once usage is linearly tied to costs, application companies must possess replaceable weights; otherwise, gross margins are determined by upstream labs. Open source weights shift the question from "can it be done" to "can the bill be converted from variable costs back to controllable capital expenditures." The top-ranked model can still command a premium, but it cannot capture all work hours.
ABAB News · Cognitive Laws
As agent usage becomes vertical, rental costs will first erode gross margins.
Open source weights are purchased for replacement rights, not for ranking first.
The most challenging tasks can still be outsourced, but high-frequency tasks must be brought back in-house.