The frontier's costs are about to be written down

Anthropic filed to go public in the same week Google shipped a multimodal model small enough to run on a laptop. Going public means explaining your costs out loud, on a schedule, forever, and the filing starts that clock on the number this industry argues about without evidence: what it costs to serve a model. Meanwhile Gemma 4 12B, Google's open multimodal model, stopped routing images and audio through separate encoders (small translator models that turn a clip into something a language model can read, at a cost in latency and memory) and fed them straight in, with quantization-aware builds aimed at phones and laptops. We said a few weeks back that renewal conversations tell you more than benchmarks do, and that still holds. Renewals just go better when both sides are reading the same numbers. Whether selling inference looks like software or like hosting, everyone has been guessing. Once they are public, the gross margin gets printed every quarter and everyone can stop.

Top Stories#

Anthropic filed to go public. The New York Times reported the filing, which sets up one of the larger listings the sector has seen. Buyers can skip past the valuation and look at the reporting calendar. A company selling inference at the top of the market will have to describe, in public and every quarter, what serving that inference costs it. Enterprises negotiating renewals have been arguing about price with no shared numbers, which is less a negotiation than a staring contest. A public reporting calendar fixes some of that, slowly, and only for the cost categories a filing chooses to break out.

Gemma 4 12B removed the separate encoders from multimodal. Google introduced the model with audio and images fed into the language model directly, rather than translated first by dedicated encoders, a change it credits with lower latency and smaller memory use. Meanwhile, Google published quantization-aware training versions, which shrink the weights without the usual quality hit, and a GGUF build for local runtimes appeared alongside them. Shipping the architecture change and the laptop-sized packaging in the same week says something about who Google pictures running this, and it is mostly someone with a laptop rather than a rented cluster.

Salesforce is buying Contentful. Salesforce signed a definitive agreement to acquire Contentful, the content platform where a company's marketing pages, docs and product copy tend to live, folding a system of record for content into a suite that already holds the customer records. Content management used to be a category you bought on its own. Now it is being absorbed into the applications that read from it. For teams building agents against company data, the practical effect is that another large pile of documents moves in under an application vendor, and becomes reachable on the vendor's terms rather than the team's.

Top Links#

• The Inference Shift (open.spotify.com): an episode on inference economics and why cheap inference is becoming a competitive difference rather than a footnote.
• A Visual Guide to Gemma 4 12B (newsletter.maartengrootendorst.com): diagrams for anyone who wants to see exactly where the encoders went.
• What is Copilot exactly? (idiallo.com): an attempt to pin down what the product actually is, which proves harder than it should be.
• AI enthusiasts are in a race against time (charitydotwtf.substack.com): an argument that the loudest camps are both reacting to real problems and mostly talking past each other.
• Elon Musk Laid Out 602 Goals. We Counted How Many He Hit. (nytimes.com): an interactive tally of stated promises against outcomes, a genre more of this industry deserves.