- Open models win where data control, cost and vendor leverage matter most, according to executives from Vast Data, Vultr, Arrcus and startup Apelogic
- Vultr CMO Kevin Cochrane predicted a 20/80 split: open models for sensitive core workloads, frontier services for everything else
- Chinese models carry the risk that Western users could be blocked from using the technology, as they were with Huawei, warns Arrcus CEO Shekar Ayyar.
Open-weight AI models are at the center of the loudest policy fight in tech, with Washington accusing China's Moonshot AI of stealing Anthropic's technology and Nvidia rallying most of the industry behind a statement defending open models.
The fight foregrounds an important question for enterprises and telcos: When does it make sense to run an open model — whose trained parameters anyone can download, inspect, fine-tune and run on their own hardware — instead of writing checks to a frontier lab such as Anthropic and OpenAI? Executives across the AI infrastructure stack told Fierce the answer comes down to data control, cost, negotiating leverage and geopolitical risk.
Data control comes first
The open-versus-closed debate misses the most important point, said John Mao, VP of global business development at Vast Data. "I don't think they care so much if it's closed weights or open weights, so long as they have assurances that their data is not being used in ways that they don't know," Mao said. Enterprises want assurances that their data isn't disseminated for profit beyond their four walls.
Open-weight models help enterprises fine-tune models using their own data, and retain the results as their own intellectual property. That's not easily done with a closed-weight model, Mao said.
Kevin Cochrane, CMO at cloud provider Vultr, agreed. He said about 20% of enterprise workloads, touching "highly privileged, confidential, proprietary data," will run on open models on infrastructure the enterprise controls. Some 80% will use prebuilt services from frontier providers.
For example, a pharmaceutical company training a model on clinical trial data will do that work on an open-weight model that it can control. "I've been in a clinical trial. You're not going to send [my data] to OpenAI. I don't consent to that," Cochrane said. "But do I want my healthcare provider to access clinical trial data in order to improve healthcare outcomes for me? Of course I do. They just need to do it in a highly secure, confidential way."
The two tiers feed each other rather than compete, Cochrane said. Open models will drive mainstream enterprise AI adoption for application-specific use cases, which then drives broader awareness and adoption of prebuilt services from providers such as OpenAI and Anthropic. "Heterogeneity always wins," Cochrane said.
The Atlantic Council, a Washington, D.C. international affairs think tank, said general-purpose AI will — counterintuitively — become a specialized need. "Almost no organization needs a generalist AI," the organization said in an analysis published Monday. "A logistics firm's model does not need to design rockets, annotate Byzantine history, or write screenplays. Fine-tuning a smaller open model on the organization's own data produces a system that outperforms a hosted generalist at the organization's actual work, fits on hardware the organization can own, and keeps the training signal ... inside the company."
As an example of the value of fine-tuned AI, AT&T last week launched OTel 2.0, a line of open source AI models optimized for telcos.
The leverage play
The endgame is commoditization, said Boris Renski, CEO of Apelogic, which builds tools for enterprises to deploy and govern open source AI agents. "The stuff you get from OpenAI and the stuff that you get from whatever neocloud that is hosting some open-weight model is going to be roughly the same, like Pepsi and Coke." That doesn't kill the frontier labs — Renski compared them to Microsoft, which wedged itself into enterprises 25 years ago and still runs the same business — but it changes what enterprises should do now.
Indeed, Nvidia CEO Jensen Huang said in March that OpenClaw is the new Linux
"Historically, open source has been an efficiency engine," Renski said. When a frontier lab's salesperson shows up to negotiate an enterprise contract, "the objective of that salesperson is going to be to extract as much margin out of you as possible, and your ability to negotiate is going to be a function of how locked in you are to a frontier lab." Even enterprises that standardize on Claude or GPT should run open models somewhere in the stack, he argued, purely for negotiating position.
Cost is why most people go to open source, and Claude "is too expensive for large enterprises" — deployments can run to hundreds of millions of dollars, said Shekar Ayyar, CEO of Arrcus. Open source keeps Anthropic and the other frontier providers in check, he said.
But Ayyar is skeptical of open source as a self-sustaining force. Open source often fails "not because of the technical ability, but because of the business," he said — it needs a known entity behind it, a Red Hat, and few companies can play that role at scale. Hyperscaler sponsorship is no guarantee either: "It's nice to be Microsoft and say I'm using SONiC," he said, but if the hyperscalers' core businesses falter, their open source patronage goes with them.
The telco role in open-weight AI
Most telcos shouldn't bother with open-weight AI, Renski said. Telcos are not innovation engines, he said — they outsource technology, buy gear from Ericsson and Nokia and win on capital efficiency and spectrum, not software differentiation. "If you're a telco, go use Copilot" or whatever your incumbent vendor supplies, he said.
The companies that should diversify into open models are those that compete on technology — financial services and e-commerce, Renski said.
Not everyone agrees with Renski about telco AI; operators like Orange and AT&T are bullish.
On the other hand, Ayyar sees a bigger telco role, as sellers of open-model infrastructure. Arrcus is building networking for distributed inference to enable companies with connected physical infrastructure — telcos, tower companies, data center operators — to become inference service providers. Telus is running a proof of concept on Arrcus' inference network fabric to deliver sovereign AI inferencing across Canadian infrastructure. Sovereignty requirements and open, self-hosted models go together: A government or bank that insists on keeping data in-country needs a model it can run in-country.
The China question
Vast's Mao cautioned enterprises and telcos not to panic about Chinese open models. "It's not like you download and suddenly all your data is being exfiltrated to China. That doesn't happen," Mao said. A downloaded model has no connection back to the lab that built it: It's "a stateless piece of software" that "just takes input and gives you output," he said — wherever it runs, the data stays there.
China is developing open models to undercut American AI leadership, Mao said. Chinese labs can't sell hosted AI services to Western enterprises — "we all know that's probably never going to happen because of all the geopolitical stuff" — so they give the models away while building strength, he said.
The big risk to using Chinese models is regulatory, Ayyar warned. A company that standardizes on a Chinese open model risks a rerun of Huawei, where Western companies were abruptly blocked from using the technology, period. That risk isn't just from the U.S.: China's Ministry of Commerce is reportedly considering export controls on model weights, which would preserve API access but cut off downloads — stranding anyone who treated Chinese open models as permanently available.
Renski counseled "lightweight regulation, primarily anchored on protecting national security" at the federal level, with enterprises otherwise free to make use-case-specific decisions. "You don't want to be like, no Chinese models are allowed by any U.S. company, period. That's too aggressive, and unnecessary."
And for a rundown of the escalating fights about open-weight AI, see our companion story, coming tomorrow.
Read more about AI infrastructure
Nvidia GTC: OpenClaw is the new Linux, says Jensen Huang
Arrcus weaves AI inference fabric PoC for Telus