Something strange is happening inside finance teams at companies that build with AI. The price they pay for a million tokens keeps dropping, sometimes dramatically. And yet the invoice at the end of the month keeps climbing. The gap between those two numbers has quietly become one of the defining cost stories of the year.
Andreessen Horowitz gave the falling-price half of that story a name in late 2024, calling it LLMflation. By its measure, the cost of running a model of a given quality has fallen roughly tenfold every year for three years. A capability that cost sixty dollars per million tokens when GPT-3 launched could be had for around six cents by the time the analysis was written. On paper, that looks like relief for anyone shipping AI features.
However, lived experience tells a different tale with total spending steadily increasing even as unit prices have collapsed, because teams consume far more inference than they used to. Token prices down sharply year over year, and aggregate AI bills rising regardless.
Agents Burn Tokens a Chatbot Never Touched
Volume drives the rising total, and most of that volume comes from a change in how AI gets used. A simple chatbot answered one question with one model call. An agent, in turn, researching a topic, drafting an output, checking it against a policy, and revising it can burn through tens or hundreds of thousands of tokens before it returns anything at all.
Cheaper units are being offset by far heavier usage, so falling prices tell only half the story. Background workloads compound the problem further. Monitoring agents, document watchers, and compliance systems run continuously, consuming compute against every event they see, whether or not a human asked for anything.
Much of the cost also sits outside the model invoice itself. Retrieval steps, retries, caching, orchestration, and observability all add tokens that rarely show up on a pricing page. The numbers teams watch keeps going down; their accountant watches keep going up.
Inference Becoming a Ceiling on Margins
Such arithmetic can decide whether a product survives, especially for a small team. Recurring inference becomes a ceiling on margins, experimentation, and on how far a feature can scale before it stops paying for itself. Noted cases of companies exhausting an annual AI budget within months once adoption among their own staff took off.
The pressure has pushed attention toward the supply side of compute. The centralized cloud remains the default for most production work, and for good reason. The hyperscalers offer dependable service levels, compliance certifications, and tightly engineered hardware that distributed alternatives still struggle to match for the most demanding jobs.
Cost and concentration are the catch: a large share of high-end GPU capacity sits with a handful of providers, and reserved instances for a sizable open model can run into tens of thousands of dollars a month. This spread between what compute costs to provide and what it costs to rent has opened room for other models to compete.
A Different Shape of Supply
One response is what observers loosely call decentralised or distributed compute. The idea is to pool GPUs that already exist, including idle gaming machines and underused capacity in small and mid-sized data centres, then route real inference work to whichever node fits best. Crop of networks pitch savings of roughly half to three quarters against hyperscaler pricing for suitable workloads.
Infrastructure watchers reach a more measured verdict than the marketing does. These networks tend to shine on batch jobs, inference, and shorter runs. For long, tightly synchronised training across continents, the latency and coordination costs still favour centralised clusters, at least for now
Established players have been moving in regardless, like Akash markets managing inference across dozens of data centres with an OpenAI-compatible interface; networks such as Hyperbolic, Render, and io.net pursue variations on the same theme. Verification has become a shared preoccupation, since a buyer routing work to strangers’ hardware needs some proof that the work was done as claimed.
The Latest Entrant From Abu Dhabi
Among the newer names is FAR Labs, the infrastructure division of the Abu Dhabi-based AI & gaming development studio Dizzaract. The company has opened builder registration for its platform, FAR AI, framing it as a distributed network that connects builder demand with available compute and routes jobs through an orchestration layer. New registrants are being offered a million free tokens to try it.
FAR Labs traces the platform back to its gaming roots, where adaptive AI systems and player economies create an appetite for flexible compute. The pitch was described as tapping idle consumer GPUs and turning them into paid infrastructure. The platform offers a single OpenAI-compatible API across multiple models, which lets developers switch providers with minimal code changes.
Security and reliability are selling points the company also leaned on, acknowledging that price alone doesn’t win production workloads. Its design includes isolated execution through trusted execution environments, encrypted communication, and a proof-of-compute mechanism. A reliability score rates nodes on uptime, job completion, latency, and incidents, so routing decisions rest on evidence rather than mere availability.
Reading the Pricing Claims
FAR Labs has shared benchmark figures with partners, claiming sizeable discounts on selected open models against named competitors. By its own account, its listed price on one Qwen variant undercut two rival providers by up to ninety-one percent, with smaller but still substantial gaps on heavier models.
Listed prices shift constantly, comparisons depend on which precision format and configuration you pick, and a figure captured on one day may not hold the next. Spheron, for instance, carries explicit warnings that GPU rates move with availability and may already be stale by the time anyone reads them.
Distributed supply can genuinely lower the cost of inference for the right kind of job, and several networks now demonstrate that in production. The size of the savings varies with the workload, the model, the latency tolerance, and the moment you happen to check.
Reliability Decides What Survives Real Users
Cost is the hook, but it rarely decides whether infrastructure survives contact with real users. Production work demands predictable uptime, low and consistent latency, data governance, and service guarantees. Distributed networks are building the verification and routing machinery to make comparable promises, and that work is ongoing rather than finished.
A hybrid conclusion many practitioners share is: send forgiving, cost-sensitive jobs to distributed supply, and keep mission-critical and tightly coupled work on traditional clouds. The two are to be treated as complementary instead of a winner-take-all contest. Reliability still varies enough that a blended strategy is the pragmatic default.
For builders, the useful posture is curiosity tempered by testing, because the economics of inference now reward teams that measure consumption early and route different jobs to different tiers. Distributed platforms like FAR AI, Akash, and their peers widen the menu of options, which is welcome.
The discipline of matching each workload to the right option is what turns that wider menu into an actual saving.
A Market Still Taking Shape
Nobody has won the inference market in 2026, and the contest is still wide open. Big cloud providers lead on reliability and sheer scale, even as distributed networks come at them on price and on tapping compute that would otherwise go unused. Underneath both camps, better hardware and more efficient models keep changing the ground they stand on.
One tension looking set to stay is that AI keeps getting cheaper to run, yet people keep consuming far more of it, and that appetite grows faster than prices drop. Whoever can meet that demand cheaply and dependably will hold real power.
Watch the pricing pages over the next year, and you will see who is pulling ahead.
