FIELD NOTES
The venture world has been feeling particularly unhinged lately.
This week, the Factory–Cognition software factory rivalry turned into a full-blown VC bitch fight. Factory accused a board adviser of sharing confidential information before joining Cognition; he denies it. Then Vinod Khosla jumped in to call Factory a “struggling second tier competitor,” while Keith Rabois (his partner at Khosla Ventures) publicly defended Factory. Khosla Ventures happens to own stakes in both companies. TechCrunch has the whole glorious mess.
Valuations are multiplying at eyewatering rates as VCs fall over themselves to get into the latest thing. TypeSafe emerged from stealth at a reported $200 million valuation; within days, investors were discussing a round above $10 billion. Instinct went from $2.5 billion to $10 billion in a month after it went viral on VC twitter, while still largely in early access.
But: the hot new entrants all have impressive copycats rolled out by both new startups and big companies within days, if not weeks.
Is there a moat? Maybe. Maybe no one cares. At the moment, the market seems perfectly happy to jump in first and figure that part out later.
-Tara
THE DOWNLOAD
Anthropic’s leaked IPO prospectus shows what it costs to run a frontier lab
Anthropic generated $4.6 billion in revenue in 2025, up from roughly $400 million the year before, according to its confidential IPO prospectus reviewed by Reuters. It also spent $7.33 billion on compute and infrastructure and posted an operating loss of just over $8 billion.
The more striking number is further out: Anthropic has committed roughly $518 billion to future cloud, compute and infrastructure spending. Most of those commitments are non-cancellable or payable regardless of actual utilization, according to the same prospectus reporting.
Its commercial relationships are equally intertwined. Nearly half of 2025 sales flowed through Amazon and Google’s cloud marketplaces — companies that are simultaneously Anthropic investors, infrastructure suppliers, distributors and competitors. Reuters reported that Amazon and Google collected hundreds of millions of dollars in distribution fees on those sales.
The reported $42 billion net loss is less useful than it looks because roughly $34 billion came from non-cash accounting charges tied to financing instruments; the operating loss gives a cleaner picture of the underlying business, as MarketWatch noted.
Why it matters: Frontier AI is beginning to look like software demand sitting on top of infrastructure economics closer to heavy industry.
OpenAI’s DevDay fills in the agent stack
OpenAI used DevDay to package more of the productization required to turn models into persistent software agents.
The new Agents API includes hosted execution, memory, tools and multi-agent orchestration. Computer use is built directly into the API. Codex tasks can keep running in the cloud.
A new Decisions API uses Luna to classify inputs, route requests or choose from predefined actions rather than generate open-ended text, according to OpenAI’s DevDay recap.
At the same time, inference is getting cheaper and faster. OpenAI says GPT-6.1 Sol costs one-fifth as much as Astra at standard token pricing, while its Ultrafast tier can increase generation speed substantially in the API and Codex, as detailed in the DevDay developer announcements.
Taken together, the architecture is becoming more modular: frontier reasoning for hard problems, smaller models for routing and decisions, plus memory, execution and tool access around them.
Why it matters: The competitive layer is moving beyond the model toward the runtime that decides when, where and how different kinds of intelligence get used.
DeepMind watermarks proteins that remain detectable after synthesis
DeepMind has demonstrated a watermark for AI-generated proteins that survives the jump from a digital sequence to a physically synthesized molecule.
SynthID Bio, published in Nature, embeds a detectable signal into generated protein sequences and predicted structures. Researchers synthesized watermarked protein binders against SARS-CoV-2 RBD, VEGF-A and PD-L1 and found that watermarking did not materially reduce binding performance.
The sequence watermark is inserted during generation by subtly changing amino-acid choices. A separate method adds a signal to predicted 3D structures using a fine-tuned AlphaFold 3 model, as described in the paper.
The obvious applications are provenance and screening: identifying AI-designed sequences entering synthesis pipelines or preventing synthetic structures from quietly contaminating biological databases.
It is not tamper-proof. Independent Nature coverage notes that the marker can be intentionally removed, and DeepMind describes the work as a proof of concept rather than a complete biosecurity system.
Why it matters: As generative models begin producing molecules rather than media, provenance becomes a physical-world problem.
Decision models are multiplying. Are they truly a new primitive or just a feature to be rolled in?
Jev launched two weeks ago with a simple premise: many jobs inside an agent do not require a language model to generate language.
A tool router, fraud check or escalation decision often comes down to choosing among a fixed set of actions. Jev calls models optimized for this decision models: they accept a state and a set of choices, then return calibrated probabilities over the allowed answers rather than free-form text. The launch quickly sparked copycats and debate over whether this is really a distinct model class, as The Wall Street Journal reported.
The idea has spread quickly.
Cloudflare released Clef and Clef-flash, open-weight decision models designed for the same class of bounded choices. Jeff offers a small Jev-compatible implementation. PostHog’s Jeeves adds a reasoning step before producing a structured decision. OpenAI’s newly announced Decisions API now handles classification, routing and action selection inside its own agent platform.
The counterexample may be the most revealing. Privatemode showed that a regular GLM model can be constrained into a Jev-like decision system in a single forward pass without purpose-training a separate model. That suggests the workflow may matter more than the architecture.
The deeper shift may simply be that agents are starting to separate open-ended reasoning from fast bounded decisions, instead of using a large autoregressive model for both.
ALSO
AMD will acquire World Labs for approximately $8.2 billion. AMD says World Labs’ model research will help shape future hardware, software and systems around emerging workloads. Fei-Fei Li will become AMD’s executive vice president and chief scientist if the deal closes, according to AMD’s announcement.
OpenAI and Synopsys are building a specialized model for chip design. GPT-Synopsys will combine OpenAI models with Synopsys’ EDA tools; OpenAI is also licensing those tools for development. Synopsys announcement
Apple is tightening Full Disk Access as agents become more autonomous. Apple says the permission can expose files, mail, messages and browsing history, and that increasingly capable agents make that access substantially riskier. Future macOS controls will require more explicit user action before granting it. Apple developer notice
Microsoft is building a multimodal world model of biology. Quine connects genomics, proteins, chemistry, cell state and imaging with scientific tools and wet-lab feedback. Researchers from Harvard Medical School and the Broad Institute participated, but Microsoft describes Quine as an early-stage research effort and there is not yet a peer-reviewed independent evaluation. Microsoft Research