AI inference cost

Salesforce AI Research and UIUC replaced the scoring pass in KV cache eviction with a uniform random draw, kept the prompt pinned, and matched the strongest published evictor across four models and six reasoning tasks — at 32-43% higher vLLM throughput. Random-plus-prompt-protection is now the baseline any 'intelligent cache compression' claim has to beat.

Qualcomm's press release called it a collaboration; its 8-K disclosed a warrant on 25,000,000 shares at $161.26 that vests as Amazon places binding purchase orders, up to $60 billion, through 2036. The silicon is already in production and AWS has announced no instance type, which makes contract language the only lever you hold this quarter.

Normalised to dollars per TB/s of memory bandwidth per hour, only AWS Trainium2 beats NVIDIA on a price you can actually get quoted. Google's best inference chip has no published rate, AMD's case dies at hyperscaler pricing, and Intel Gaudi 3 has no successor.

Anthropic cut Claude Fable 5.1 cache reads 75% to $0.25/MTok, half of Opus 5's rate on a model with twice the base input price. But Opus 5 stays cheaper until cache reads pass a third of your bill, and Anthropic's own savings ceiling of 45% falls short of the 50% crossover.

Tencent's Hy4 preview markets 49B active parameters, but all 770B must stay resident — roughly 780GB in FP8, which an 8xH100 node cannot load at all. Size a Mixture-of-Experts model on total parameters and price throughput on active ones.

Claudeforce puts 37 prebuilt sales skills inside Claude, with open beta in September 2026 and no published price. Salesforce meters MCP tool calls against your org's shared daily API allocation; Anthropic bills inference on a separate contract with no ceiling.

OpenAI cut GPT-5.6 Sol to $4/$20 per million tokens but guarantees the rate only "at least through November 21, 2026," with no successor published. The cut was asymmetric, so no two workloads saved the same amount — and the reversion is +50% on output, not +33%.

Anthropic is in advanced talks to buy Decart for roughly $7 billion, mostly in its own pre-IPO stock, to cut what a Claude token costs it to produce. Every Sonnet through 4.6 still lists at its March 2024 price, and the one cut that did land arrived as a new model number — which is why a flat multi-year rate card is the wrong thing to sign this quarter.

DeepSeek warned of a significant API price rise with no number and no date. Because the weights are MIT-licensed, the ceiling is already public: independent hosts serving the identical model charge 3-4x on posted rates and 40x on cache hits.

Databricks lists 47 servable models to Snowflake's ~30, but Unity Catalog keeps Delta tables read-only to outside engines and Snowflake's own catalog blocks third-party writes outright. Snowflake's exit advantage is narrower than the decks claim: it writes into a catalog you own. On a normalised 10-million-ticket workload the platforms land within 2x of each other; the model tier swings the bill 40x.

North Carolina's 2026 budget repealed the sales-tax exemption on data center electricity and left the exemptions on servers and construction materials intact. The 7% now falls on the meter — which prices always-on inference, not the build.