Calix Plans Small, Domain-Specific Language Models Running on Subscriber and Network Hardware, and Promises a 2027 Product Surge After a Slow Delivery Year
Calix ConneXions Analyst/Investor Day, Las Vegas, October 19, 2026
The most consequential disclosure at Calix's customer and investor event was a change in where its AI stack will run. CEO Michael Weening announced the CLM Foundation, short for Communications Language Model Foundation. It is a program to distill large open-weight models into small, broadband-specific models that Calix calls Gems. Weening said the models are not meant for the cloud. They are meant for the edge, on the company's own silicon. "The future of AI is at a local level on the edge, and we've got all that processing power just waiting for us to leverage together," he said. The company has built what it describes as repeatable processes for shrinking 30 billion-parameter models to 3 billion, 2 billion, 1 billion and even 100 million parameters. They will be sized from extra small to large and organized by function, including marketing, security, repair and troubleshooting.
Weening gave concrete deployment examples. An extra-small cybersecurity model would run on a GigaSpire in the subscriber's home or business. A medium model would run on a managed network product to protect business subscribers. He said CLM Foundation is the first of "10, 15, 20" such foundations. Calix plans to deliver the Gems through normal release cycles and a consumption-based enterprise agreement, so operators do not have to manage models themselves. The strategy rests on Weening's view that open-weight models from Google, Meta, Mistral, Microsoft and NVIDIA will keep leapfrogging each other. The architecture is designed so that Calix can "rip the brain out" and swap in a different model. He argued that proprietary labs are structurally disadvantaged because the open-weight developers do not need to earn a profit on AI. That is a management opinion, not a demonstrated outcome. Calix gave no timeline, accuracy benchmarks or pricing for the Gems.
A Slow 2025 Acknowledged, and a Large 2027 Commitment
Weening was unusually direct about execution. Product output peaked in 2024, and 2025 was "a slow year for putting out product that you can consume." The company was finishing its third-generation platform, which launched in March 2026. He said implementations slowed and that operators absorbed real disruption. "We know we slowed down some implementations," he said, and he called the effort "hard, but it was worthwhile." Every customer is now AI-enabled on the platform, though only a couple of hundred operators have adopted it end-to-end as first movers.
Management is promising a sharp reversal. Product chief Shane Eleniak has committed to the board that 2027 will deliver more capability across AI, PON and automation than the previous two years combined. Weening cited a figure of roughly 1,400 deliverables for next year, although the number was stated loosely. Two factors underpin the claim: the platform is now in place, and AI is compressing internal development. One feature scheduled for 15 man-months was reportedly completed in 3. Calix has about 900 engineers. It is spending almost $200 million on the platform this year, on top of more than $2 billion invested over 15 years. Weening said the company will consume AI from Salesforce, ServiceNow and Oracle for its own operations rather than build it. He also said Calix will not "chase the shiny nickel" of data centers. The commitment is credible only if it is delivered, and no revenue or margin guidance was offered at the event.
Workflow Studio and the November Agent Releases
Calix announced Workflow Studio, a tool that lets operators customize agent workflows built from reference templates drawn from the practices of more than 1,000 customers. The tool draws on agent-to-agent and Model Context Protocol connections to partners such as Salesforce, ServiceNow, NISC, Gaiia and HubSpot. Reference workflows for operations, service, marketing and subscriber experience are due in November. CommandIQ 4, which is subscriber-first and agent-driven, is ready for release. Chief Marketing Officer Juergen Lindner said the agent-to-agent standard was agreed in about 10 weeks. He compared that with the roughly 2.5 years he said earlier integration standards required. Calix has also built an MCP server that can sit in front of slower-moving partners.
The operator testimony was specific. Allo CEO Shane West said an operations agent cut the diagnosis of an intermittent connectivity complaint from 15 minutes to 30 seconds. At Conexon, John Walburn said outdoor WiFi attach is 0.72% of the installed base and about 2.1% on new subscribers. A segmented email campaign built with the marketing agent produced about a 2x response, which Weening said implies roughly 4%. CoastConnect reported an NPS of 91, 99.9% of issues resolved within 24 hours and 70% app usage. It plans to use the subscriber agent to protect those results without adding headcount.
The Simplification Pitch and Network Economics
Weening continued to push a platform-consolidation message. He urged operators to eliminate Nokia and Juniper BNGs and separate Cisco PE routers in favor of the E9-2, E7-2 and E3-2 stack running on one operating system. "If you're running a Nokia BNG or a Juniper BNG in it, you are disadvantaged," he said. He said Verizon re-ran a large RFP and chose Calix again. He also cited Verizon's quoted 80% lower cost to run a network compared with the legacy architecture. He named Brightspeed as another customer. Allo's data supported the operating-cost argument. West said Allo runs a 25-person IT staff against 400 at a multi-vendor competitor. Its trouble ticket index is 40% lower, and its NPS is about 60 versus roughly 0 at the multi-platform network he compared it with. Those are operator-provided comparisons, not audited data.
The hardware roadmap is broadening. Calix announced a managed switch portfolio, including small units for homes and larger ones for multi-dwelling units. It also announced a WiFi 7 outdoor unit that begins shipping in February, WiFi 8 leadership, 50-gig PON, hospitality offerings and grid monitoring with SCADA capabilities for utility customers. Calix is also rolling out management of ADTRAN and Nokia equipment through its Operations Cloud, which acknowledges that mixed-vendor networks will persist during transitions.
Penetration Gaps That Frame the Revenue Opportunity
Weening was blunt about adoption of existing products. Calix estimates a 12% target for outdoor WiFi, based on 8% of US homes with pools and 4% with recreational ponds. The attach rate across the installed GigaCenter and GigaSpire base is 0.5%. "You might as well burn a $10 bill every single month," he said. SmartTown is switched on by only 16% of operators' subscribers, against what he called an obvious 80% to 100% target, even though it is free. He said a Calix survey found 75% of consumers want a community WiFi experience. Weening also named Amazon's eero as the industry's biggest competitive threat because it reduces fiber providers to "a dumb pipe." He said the shift would make fiber providers interchangeable. Operators are also facing higher memory costs, which he said have risen about 5x, and rising interest rates. He predicted that the AI-driven memory price spike "will bust."
Security and Predictability as the Selling Points
Weening spent considerable time on trust, arguing that the data, knowledge, identity and orchestration layers allow the company to treat AI agents as tightly permissioned employees. He recounted an episode in which a customer tested the Calix agent by repeatedly asking it how many calories are in a McDouble. After 51 repetitions the agent answered "390 calories." It did not expose customer data or execute actions, he said, and Calix patched the weakness. The admission is useful context. The failure was a guardrail breach, and it indicates how much adversarial testing remains. Weening also positioned cost predictability as a differentiator, pointing to reports of enterprises blowing through AI token budgets. He said his own board's cybersecurity review has grown from 30 minutes to 3 hours. He also said AI cut production costs for event videos from $800,000 to $1 million to a team of 3 or 4 people.
The event was a customer conference, not a financial update. There was no quantified guidance and no AI revenue disclosure. The substantive new information was the shift toward distilled edge models, the November workflow releases and a delivery commitment for 2027. Investors will want to see whether product delivery returns to its 2024 pace and whether attach rates for outdoor WiFi, SmartTown and agent products move from anecdotes to figures that appear in the company's results.