The great AI slowdown has begun. Could it be an opportunity for telcos to get back in the game, asks Stewart Baines?
OpenAI has put off going public while it works on model safety, Anthropic is investigating the welfare of its models and wants to slow down model development. And the hyperscalers and neoclouds building AI factories are discovering that their construction schedules are being rewritten by protestors, planners and power constraints. Hundreds of billions have been spent on GPUs with nowhere to go.
From the demand side, many enterprises are regretting tokenmaxxing and now want AI to pay its way. But tracking the bill is difficult when input, output and cached tokens carry different prices. New models can mean prices change, and labs “silently change limits whenever they want” reports SemiAnalysis. Respondents to the FinOps Foundation’s 2026 survey rank AI cost management as the skill they most need to develop.
None of this seems to be deterring frontier labs pouring money into model development, judging by Microsoft AI chief Mustafa Suleyman’s forecast on The Rest Is Politics: Leading. He predicted training runs costing “many, many tens of billions of dollars, potentially reaching US$100 billion, within a couple of years, with only five or six labs able to afford them.”
The prize for frontier labs and hyperscalers is control of a technology that they believe is the most important ever. Governments around the world are torn: they know they need to support AI development to be competitive, yet they are constrained by the need to strengthen national resilience and enforce their own rules. They have good reason to worry about ceding control to a handful of US companies, and about AI access depending on the whims of the White House.
Asked if it’s game over for AI competitors, Suleyman believes there is a route to independence: sovereign data centres running models that cannot be withdrawn at a foreign government’s request. Whether open source or open weight language models, they could be trained on data such as defence or health which is critical to national resilience.
So here lies an opening for telcos and their data centre partners, whose fibre, facilities and local operations could help enterprises and public sector organisations build AI services with their own data that they can keep under their own control, choosing where the processing takes place and who operates the service.
Local hero versus the 800lb gorillas
Enterprises that prioritise network sovereignty rank telcos as their most trusted partners for AI investment, according to Omdia which means that operators with national fibre footprints and local teams have a distinct advantage over hyperscalers and systems integrators. However, demonstrating that the trust is not misplaced means being able to show where the AI data travels, who can access it and how service continues if a supplier or connection fails.
To create a sovereign AI infrastructure, we don’t need to wait years for greenfield gigawatt AI factories to come on stream. Some of the capacity could come from knitting together retrofitted brownfield facilities into a connected pool of AI compute. EY-Parthenon’s February 2026 report, Hyperscale meets hyperlocal, identifies grid infrastructure, land and permitting as constraints in established data centre hubs, giving existing sites with spare power capacity an advantage. NVIDIA-backed Nscale’s 90MW data centre in Loughton was due to come on stream in 2027 but reportedly faces a wait into the 2030s for sufficient grid power.
Richard Vosper, chief revenue officer at fibre network operator Zayo, sees business beyond the giant campuses: “I don’t expect inference to sit entirely within hyperscaler clouds.” He advocates upgrading smaller and regional facilities to bring inference closer to users and their data.
The appeal is that much of the expensive work has already been done: existing sites have grid connections, fibre routes and carrier interconnections. Reusing those assets could get capacity into service without starting again with land, ducts and civil works. An existing grid connection does not guarantee spare power, however, and a building designed for conventional server racks may need substantial changes to accommodate denser power-hungry AI hardware.
UK data centre and cloud infrastructure operator Pulsant has been doing that at Milton Keynes, where CMO Mark Lewis says: “New cooling and power delivery systems mean that the facility can now run both traditional density and high-density workloads in the same building.” The challenge is to accommodate those loads while keeping power and cooling costs under control. Better cooling can reduce overheads, but it does not make the accelerators themselves less hungry, and the investment still needs enough paying customers to justify it.
Place the workload where it makes sense
If the goal is to process documents, spot fraud or inspect production lines, enterprises don’t necessarily need the latest frontier model from Anthropic or OpenAI. Smaller, specialised models may meet those needs with less compute, and some tasks need no language model at all. Downloadable open-weight models, like Microsoft’s Phi family, give enterprises more choice over hosting, including within sovereign infrastructure. Some proprietary models can also be offered locally under specific deployment arrangements but read the fine print.
Large training runs required for frontier models generally favour tightly coupled accelerators in one location, whereas inference can use copies of a model at different sites to serve requests wherever the economics and application requirements allow. A chatty application may need compute close to users while overnight processing has more freedom to follow cheaper capacity. Each request can still be processed within one facility.
Richard Petrie, chief technology officer at the London Internet Exchange, argues that customers should not have to worry about locating every resource: “Instead, they could have access to a pool of computing resources across different locations, with the network helping to connect them to the most appropriate capacity at any given time.”
Choosing that capacity means checking which sites have the required model and data, where the regulations or enterprises permit processing, and what the combined compute and network bill will be. The nearest GPU might have a queue long enough to wipe out the latency advantage, while the cheapest capacity may come with a hefty data-transfer bill.
Connecting those sites into a pool of AI capacity requires sufficient bandwidth for model distribution, data transfers and application traffic. Coherent optics from vendors like Cisco and Ciena can help operators carry more data over existing fibre and connect facilities hundreds of kilometres apart, expanding the choice of location. Distance still introduces latency, but deep-buffer routers can absorb traffic bursts and reduce packet loss, helping maintain throughput over longer links. Diverse routing can improve resilience, while traffic management can give interactive inference priority over less urgent transfers.
BT’s InferSlice demonstration with NVIDIA and JOINER at Connected Britain showed how network and compute resources could be coordinated, allocating them to different service tiers for distributed inference. The idea is to route requests according to GPU capacity while managing different classes of traffic end to end. Customers could bring their own models and the demo also divided GPU capacity between tiers. That suggests a way to sell different levels of service while sharing costly accelerators, provided operators can maintain isolation and predictable performance.
A service worth paying for
Telcos will have competition from CDN providers such as Akamai and Cloudflare. Akamai’s AI Grid orchestrator routes inference across edge, regional and core infrastructure to balance cost and performance, while Cloudflare’s Workers AI provides inference through an API across its network. Operators could partner with these platforms, but simply owning the fibre will not guarantee them the customer relationship.
An operator and its data centre partners could sell managed interconnection, reserved compute and location controls as one service, with capacity for an interactive customer-service application in one facility and overnight document processing at another approved site. Usage reporting and cost limits could keep spend in check, while performance commitments would need to specify the model, request size and concurrency covered. The operator may also need partners capable of running the compute and model, and agreeing who is responsible for security, updates and support.
Matt Partridge, managing director for enterprise and wholesale at LightSpeed Networks, identifies what customers should expect for their money: “For service assurance, it means proactive monitoring, clear accountability and a single point of ownership when something goes wrong.”
Enterprises buying an AI service will have little patience for disputes between the network operator and its hosting partners, so those partners need to agree on who monitors the complete service and fixes failures. Telcos can earn a share of the AI bill without financing the next frontier training run, provided they can turn their infrastructure into a service enterprises can afford to rely on.
Stewart Baines