Cloudflare entered the inference pricing race with a multimodal decision model at $0.038 per million tokens. Cloudflare released Clef-omni, built on a 30-billion-parameter base with a 3-billion active parameter mixture-of-experts architecture, licensed Apache 2.0. It scores text, images, audio, and video in a single API call. The standard tier runs $0.15 per million input tokens; the faster Clef-flash variant drops to $0.038. That flash price is aggressive even against bare-bones text-only endpoints from most providers. (MarkTechPost)
Brief
Monday, October 12, 2026 · 4 min read
Microsoft's Maia 300 chip targets Nvidia inference dependency
Good morning. Microsoft is on its third in-house AI chip, Cloudflare is in the inference pricing business, and a rogues' gallery of open-weight decision models dropped over the weekend. The week opened with more infrastructure news than most months used to produce. Somewhere, a Nvidia account manager is having a difficult Monday.
Today's reading time is 4 minutes.
Listen · Snippt Daily0:00 / 4:04
INDUSTRY
Every major cloud provider has been running AI workloads almost entirely on Nvidia silicon, which means Nvidia sets the effective floor on what inference costs.
Driving the news: Microsoft announced Maia 300, its third-generation in-house AI chip, designed to serve as the 'primary AI inference system' for its own products and services. The chip is aimed squarely at reducing what Microsoft pays to run models in its data centers, not at training new ones. It is rolling out now inside Microsoft's cloud infrastructure. No external pricing was disclosed, but the framing as a primary inference system suggests it is meant to carry real production load, not sit in a pilot rack.
- Maia 300 is the third generation of Microsoft's in-house AI silicon, following Maia 100.
- The chip targets inference specifically - running models at scale - rather than the training workloads Nvidia's H100 and B200 dominate.
Zoom in: Microsoft has been buying Nvidia GPUs at enormous scale to power Azure AI services and its own Copilot products. Custom silicon is the standard move for hyperscalers once their workloads are large enough that shaving a few cents per inference token adds up to hundreds of millions of dollars annually. Google's TPUs and Amazon's Trainium and Inferentia chips followed the same logic. Microsoft's first Maia chip arrived in 2023; getting to a third generation in under three years suggests the program is being taken seriously internally.
- Inference, not training, is now the dominant cost for deployed AI products - which is exactly what Maia 300 targets.
- Microsoft's Copilot suite runs billions of inference calls daily, making even small per-call savings material at that volume.
Why it matters: For builders using Azure, this is mostly invisible in the short term - you call an API, the chip underneath is Microsoft's problem. The longer-term effect is that Microsoft has more room to cut inference pricing without waiting for Nvidia to move first. That matters most for anyone running high-volume workloads where token costs are the dominant line item.
Bottom line: Microsoft is not leaving Nvidia behind, but it is building the leverage to stop asking permission before dropping prices.
The Verge ↗Get this every weekday
Two minutes, 7am ET. No fluff.
Also happening
vLLM Semantic Router shipped Decision 3.0, five open-weight models that classify without generating text. Decision 3.0 covers five model sizes from 0.8B to 27B parameters, all Apache 2.0 licensed. They handle yes/no, multiple-choice, and scoring tasks over text and images in one call, skipping the token generation step that makes standard LLMs slow and expensive for routing and moderation tasks. Skipping generation is the key design choice: it makes these models faster and cheaper for high-volume classification pipelines than a general-purpose model asked to reason through an answer. (MarkTechPost)
OrcaRouter released a cybersecurity model claiming 100% on Cybench with a 1-million-token context window. OrcaCyber Zero 1.5 is a gated model - you apply for access rather than pulling it freely - priced at $3.00 per million input tokens and $7.50 per million output tokens. The 1M-token context window is the headline spec, useful for ingesting large codebases or log files in a single pass. It reports 95.8% on an evaluable subset of CVE-Bench, a benchmark covering real-world vulnerability exploitation tasks. Gated access and specialty pricing put it squarely in the enterprise security buyer market, not the hobbyist tier. (MarkTechPost)
The thread
Inference cost is the new battleground
Three of today's stories are, at root, about the same thing: who controls the price of running a model call. Microsoft is building its own silicon to escape Nvidia's pricing gravity on inference. Cloudflare is offering multimodal classification at $0.038 per million tokens. vLLM Semantic Router is shipping open-weight models that skip text generation entirely to cut per-call cost. For anyone building products where inference volume is the main expense, the options and the leverage are both expanding.
From the publisher
Snippt
Image, video and voice tools in one place
On our radar
- SpireBench is a new benchmark using the roguelike card game Slay the Spire to test whether agentic models can reason strategically over hundreds of sequential decisions, not just short episodic tasks.
- EliteTable is a self-improving program-search algorithm where an LLM maintains a fixed table of high-scoring solutions and iteratively edits or remixes them, showing gains on control policy tasks.
- A new arXiv paper argues neural networks are more modular than individual neuron analysis suggests, introducing a formal framework that finds structure at the level of circuits rather than single units.
- Voice AI executives told TechCrunch the technology still breaks too often at the context layer - the part that tracks what was said earlier in a conversation - to have its mass-adoption moment.
- A Verge writer documented the experience of setting up local AI models on a Mac Studio to avoid sending personal data to cloud services, finding it exciting, overwhelming, and frequently frustrating in roughly equal measure.
- TLDR AI flagged Nvidia weighing an acquihire of Reflection AI and a Gemini 4 Carbon model reference, neither confirmed by primary sources as of this morning.
AI worth acting on, in your inbox
Get AI Weekly every Friday: the AI news, tools and tactics that matter for marketers, sellers and creators. Two minutes, with audio.
You can change these choices or unsubscribe at any time. Privacy Policy