Inference Wars @InferenceWars

Week #68: AI is splitting into decision + execution. Cheap models decide: route, retry, escalate, stop. Expensive models do the hard work. The emerging stack is: intent → policy → decision → allocation → execution → verification The model is no longer the whole system.

Inference War #66 🛰️ A 100x price gap now separates GPT-6 Luna and Astra. That makes “which model?” a question worth asking at every stage of a task. Cheap worker. Stronger model when needed. Verify the result. The opportunity: knowing when to spend more intelligence. #inference #inferencewars

Inference War #65 🛰️ A GPU without power is idle. An agent without permission is stuck. Cheap tokens that need three retries can be expensive. The next inference battle: turning available compute into verified work. Watch cost per completed task. InferenceWars.com

The inference war isn’t just about cheaper tokens anymore. It’s about who controls the job: models, tools, permissions, memory and recovery. More choice underneath. More dependence on whoever runs the system. Week #64: war report, API surveillance + leaderboard ↓ InferenceWars.com

The agent is becoming the buyer of inference…and that changes who controls demand. Claude Fable 5.1, Gemini 3.8 Flash and GPT-6 Astra aren’t just competing on model capability. They’re turning agent sessions into stateful, permissioned buyers that decide when to call another model, use a tool, retry, escalate or spend more reasoning. This week’s report follows the battle from capability entitlements and session reliability to Nvidia, Broadcom, Microsoft, Vertiv, Lambda and the infrastructure needed to serve demand that cannot be predicted when a task begins. Who wins when the runtime (*not the user) controls the next inference purchase? Read Inference War + Spy Report #63: “The Agent Becomes the Buyer” InferenceWars.com

🛰️ INFERENCE WAR - WEEK #61 Something interesting is happening. The infrastructure required to produce intelligence is getting more expensive. But intelligence itself is getting cheaper. This week: • Nvidia backed up to $105B of OpenAI data-centre commitments • Google deepened its custom-silicon push with Marvell • Cerebras launched inference-focused CS-4 • CoreWeave took Vera Rubin into quantitative finance with HRT • Anthropic pushed computer use, browser use, files + skills into production • OpenAI cut GPT-5.6 Sol API pricing by 20%+ The flywheel: more capital → more compute → more competition → cheaper inference → more viable workloads → more demand The next war isn’t simply over the best model or most GPUs. It’s over who can turn power into useful intelligence at the lowest cost — and route every workload to it. #InferenceWars #AIInfra #Inference #GPUs #AIInfrastructure #Agents

🛰️ INFERENCE WAR - WEEK #60 Inference is becoming an institutional asset class. This week: • Nvidia launched financing partnerships targeting $500B+ of AI infrastructure capital • CoreWeave backlog hit $104.2B and near-term capacity remains effectively sold out • IBM + Together AI committed $240M to dedicated open-weight inference capacity • Nvidia launched NeMo Switchyard for multi-model routing • Meta pushed further into local open-weight inference • Anthropic’s latest agent controls added budgets, advisor models and inference geography • Microsoft continued pushing Maia deeper into the custom-silicon race The bigger shift: The market is fragmenting across models, clouds and chips. That makes the layer sitting above them more valuable. application → router → model → silicon → capacity The next strategic battleground may not be who owns the best model. It may be who decides where every inference request runs. #InferenceWars #AIInfra #Inference #GPUs #AIInfrastructure #Cloud #OpenSourceAI

🛰️ INFERENCE WAR - WEEK #59 The AI infrastructure race just moved from financing risk to utilisation risk. This week: • Big Tech’s future data-centre lease commitments passed $1T • Texas paused new data-centre grid approvals • AMD bought Taalas to deepen its inference-specific silicon stack • OpenAI launched faster long-context inference + API-key cost attribution • Anthropic moved enterprise policy directly into the pre-inference path • Frontier AI safety testing split between closed and open-weight models The key shift: The question is no longer: “Can we finance enough compute?” It’s: Can we keep all this committed capacity profitably utilised? The next winners will optimise: utilisation → routing → latency → cost → policy → reliability Inference is becoming an allocation problem. #InferenceWars #AIInfra #Inference #Datacenters #GPUs #Cloud #AIInfrastructure

🛰️ INFERENCE WARS - WEEK #58 The AI infrastructure war has entered a new phase: Credit is becoming part of the compute stack. This week: • Nvidia is reportedly considering backing up to $250B in financing for an OpenAI-linked data centre • Meta and BlackRock formed a $14B AI infrastructure JV • Anthropic is reportedly linked to a $15B Google-backed data-centre financing • AMD secured up to 2.5 GW of AI capacity through Core Scientific • The EU committed €10B toward seven AI gigafactories • OpenAI cut API prices while adding premium low-latency “Fast” mode The new competitive moat isn’t just chips. It’s the ability to finance: GPUs → power → data centres → long-term capacity → future inference demand. The winners won’t simply build the most compute. They’ll make the next generation of AI infrastructure bankable. #InferenceWars #AIInfra #Datacenters #GPUs #AIInfrastructure #Inference #CloudComputing

🛰️ INFERENCE WAR - WEEK #57 The inference market is moving into its pre-financed capacity phase. Key signals: • AMD agreed to supply Anthropic with up to 2 GW of infrastructure and invest up to $5B • OpenAI committed to deploying AMD’s Helios rack-scale systems • Nvidia and SK Group linked accelerators, HBM and a planned 2 GW data centre • Server CPUs and optical interconnects emerged as secondary bottlenecks • OpenAI added project-level spending controls • Anthropic expanded agent memory, lifecycle hooks and effort controls • Google launched cheaper Flash models for high-volume agent workloads The shift is clear: AI infrastructure is no longer being built after demand appears. It is being financed, reserved and integrated years in advance. The new competition is not just for GPUs. It is for the complete stack: chips → memory → racks → power → data centres → model demand #InferenceWars #AIInfra #Inference #GPUs #DataCenters #AIInfrastructure

🛰️ INFERENCE WAR - WEEK #56 The battlefield moved again. This week was not about the best model. It was about permission. Key signals: • China promoted an alternative stack built on open models, domestic chips and sovereign infrastructure • The U.S. tightened chip access through a permissioned export regime • New York imposed a moratorium on large new data centres • The White House pushed AI companies to fund the grid expansion they require • The EU forced Google to open parts of Android and search data to rival assistants • Tower Semiconductor expanded silicon-photonics capacity • OpenAI showed increasingly fragmented reliability across Codex, Sora, voice, files and enterprise services The new question is no longer: “Which model is fastest?” It’s: Is the model available, legal, powered, connected, distributed, reliable and permitted for this workload? Inference is becoming permissioned infrastructure. #InferenceWars #AIInfra #Inference #Compute #SovereignAI #Datacenters #AIInfrastructure

🛰️ INFERENCE WAR - WEEK #55 Meta launched paid model access while preparing custom silicon. DeepSeek is doing the same. Anthropic tightened API governance. Google added agent observability. The shift: model companies are becoming integrated inference utilities. #InferenceWars

🛰️ INFERENCE WAR - WEEK #54 This week exposed a critical truth: AI capacity is abundant in aggregate, but scarce in the exact form buyers need. Key signals: • Google reportedly capped Meta’s Gemini usage • Meta is reportedly exploring selling excess AI compute • Anthropic launched Sonnet 5 with 1M context and altered token economics • OpenAI showed Codex quota pressure • Gemini expanded into low-latency image and video inference • China advanced a domestic model, chip and memory stack • Neoclouds faced the risk of major customers becoming competitors The old question was: “Who owns the most GPUs?” The new question is: Who can match the right model, silicon, memory, region, quota, reliability and cost profile to each workload? Inference is becoming a market for non-fungible capacity. #InferenceWars #AIInfra #Inference #Compute #GPUs #Cloud #AIInfrastructure

Week #53: The inference war entered the full-stack control phase. OpenAI launched a custom chip (Jalapeño, with Broadcom). Qualcomm paid $3.9B for hardware-neutral inference software. Anthropic locked in a memory + storage deal with Micron. Cerebras posted its first earnings - strong demand, weak margins. The question is no longer who has the best model. It's who controls the silicon, memory, runtime, and application layer - or can abstract across all of them. Full briefing + leaderboard: inferencewars.com #InferenceWars #AI #AIInfrastructure

🛰️ Full Week #52 briefing — leaderboard, API surveillance, operator playbook, and intelligence report: inferencewars.com #InferenceWars #AI #LLMs #GPUCompute #AIInfrastructure

One year ago this tracked: → GPU $/hr Now it tracks: → Export controls on frontier models → G7 trusted-partner access frameworks → $35B+ private-credit compute structures → Nvidia as an infrastructure bond issuer Same war. Completely different battlefield.

🎂 Week 52. One year of Inference Wars. 52 weeks ago this started as a GPU spot-price board. This week it tracked sovereign access controls, G7 trust frameworks, and compliance-gated compute. The war keeps escalating. So does the coverage. 🧵