STEVEN STEINBECK
← ALL DISPATCHESDISPATCH 41

WEEK OF OCT 4 – 10, 2025

OpenAI turned ChatGPT into a runtime, AMD bet on it in its own shares, and Ant open-sourced a trillion parameters

OpenAI spent DevDay turning ChatGPT into a software platform, AMD agreed to sell it six gigawatts of silicon in exchange for equity in itself, the MPA told OpenAI that policing Sora 2 is OpenAI's job, and Ant Group put a trillion-parameter MoE on Hugging Face under MIT.

OpenAI spent Monday recasting ChatGPT as a software platform: an Apps SDK built as an extension of the Model Context Protocol, an AgentKit stack covering agent building, chat UI, connectors and evals, and Codex out of beta. Hours later AMD agreed to supply six gigawatts of Instinct GPUs while taking a large part of its upside in a warrant for its own shares.

The same Monday, the Motion Picture Association told OpenAI that preventing Sora 2 infringement is OpenAI's responsibility, not rightsholders' — the day OpenAI put Sora 2 into the developer API. Meanwhile the open-weights line kept compressing: Ant Group's inclusionAI published a trillion-parameter mixture-of-experts model under MIT, and Google gave developers a computer-use model that treats the browser as the API.

The through-line is where value is moving: to the host runtime that runs apps and agents, to second-source compute bought with equity, and to permissively licensed weights that reset the cost floor. Model launches, funding rounds and per-seat pricing are downstream of those three.

THU · Oct 9, 2025open weightsmoeant groupinference cost

Ant Group open-sources Ling-1T: 1T parameters, about 50B active, MIT license

Ant Group's inclusionAI published and open-sourced Ling-1T on 9 October: a trillion-parameter mixture-of-experts model that activates roughly 50 billion parameters per token at a 1/32 routing ratio, released under the MIT license. Context is 32K, extensible to 128K with YaRN, and it was pre-trained on more than 20 trillion tokens. The team says it is the largest FP8-trained foundation model to date, with FP8 mixed precision giving a 15%-plus end-to-end training speedup over BF16 at under 0.1% loss deviation, and reports 70.42% on AIME 2025 at around 4,000 output tokens per problem plus roughly 70% tool-call accuracy on BFCL V3 with only light instruction tuning.

It follows Ring-1T-preview, the trillion-parameter thinking model Ant open-sourced in September, and sits at the top of a three-line family — Ling non-thinking, Ring thinking, Ming multimodal. Weights are on Hugging Face and ModelScope, where the repository lists about 999.71 billion safetensors parameters and roughly 2 TB of weights; the maker's serving guidance assumes a multi-node cluster rather than a single box.

Ant Group open-sources Ling-1T: 1T parameters, about 50B active, MIT license
Hugging Face

WHY IT MATTERS

Sparse activation is the entire cost argument: Ling-1T carries trillion-parameter capacity at something like 50B-dense latency, and MIT means no per-token license fee — your marginal cost is cluster access, not vendor inference pricing. For data-residency or air-gapped workloads that is now a frontier-adjacent model you can run inside your own VPC, paid for in multi-node serving and about 2 TB of weights to move and store.

Ant Group: Ling AI model family and trillion-parameter Ling-1TLing-1T model card (Hugging Face, inclusionAI)
TUE · Oct 7, 2025google deepmindagentscomputer usegemini

Google ships a computer-use model, and the browser becomes the API

Google DeepMind released Gemini 2.5 Computer Use in preview on 7 October, available through the Gemini API in Google AI Studio and Vertex AI. It is a specialized version of Gemini 2.5 Pro post-trained for UI control, and the capability is exposed as a new computer_use tool meant to run in a loop: the user's request, a screenshot and a history of recent actions go in, and function calls — click, type, scroll, navigate a dropdown — come out. Google claims it beats leading alternatives across several web and mobile control benchmarks at lower latency, with evaluations run in-house and by Browserbase, whose headless browsers host Google's public demos.

The model card is unusually specific about scope: it is optimized for web browsers, is not yet tuned for OS-level control, shows only promise on mobile, and requires explicit end-user confirmation for certain actions such as making a purchase. Two days later, on 9 October, Google Cloud launched Gemini Enterprise, an agent platform priced from $21 to $30 per seat per month that wires in the Agent2Agent protocol, the Model Context Protocol and the Agent Payments Protocol.

Google ships a computer-use model, and the browser becomes the API
storage.googleapis.com

WHY IT MATTERS

Treating pixels as an interface contract makes the long tail of enterprise and legacy software with no API addressable, which changes the economics of RPA and makes UI regression testing automatable inside CI. Budget for what it adds: screenshot tokens and multi-step latency in the loop, a human confirmation gate on anything that moves money, and the fact that browser sessions, logins and authenticated state are now part of your agent's attack surface.

Google: introducing the Gemini 2.5 Computer Use modelGemini 2.5 Computer Use model card (PDF)Google Cloud: introducing Gemini Enterprise
MON · Oct 6, 2025openaiagentsmcpdeveloper platforms

DevDay: ChatGPT becomes an app platform, and the Apps SDK rides on MCP

At DevDay in San Francisco on 6 October, OpenAI opened the Apps SDK in preview: third-party software can render interactive interfaces inside ChatGPT, with context shared through the Model Context Protocol. Pilot partners shipped the same day — Booking.com, Canva, Coursera, Expedia, Figma, Spotify and Zillow — and OpenAI framed the reach as over 800 million ChatGPT users. Apps went live to logged-in Free, Go, Plus and Pro users outside the EEA, Switzerland and the UK, with app submissions and a directory promised later this year.

The same keynote delivered AgentKit: a visual Agent Builder in beta, the embeddable ChatKit UI at general availability, an expanded Evals system with trace grading and prompt optimization, and a Connector Registry in beta for governing MCP servers across ChatGPT, Enterprise and the API. Codex reached general availability, GPT-5 Pro and Sora 2 entered the API, and gpt-realtime-mini shipped alongside. OpenAI says monetization will run through the Agentic Commerce Protocol, its open standard for checkout inside the chat.

WHY IT MATTERS

Orchestration, guardrails, evals and versioning are precisely the layers teams hand-roll on top of a model API, and OpenAI just bundled them at standard API model pricing — fast to ship on, harder to leave. Because the Apps SDK is MCP-native, your MCP server becomes the integration surface for the ChatGPT host and for other clients, so the lock-in worth tracking is the host layer, not the API; OpenAI's promise to publish an open implementation of the host environment is the item that would defuse it.

OpenAI: introducing apps in ChatGPT and the new Apps SDKApps SDK documentation
MON · Oct 6, 2025amdopenaicomputehardware

AMD pays in its own stock: 6 GW of Instinct for OpenAI, warrants for up to 160M shares

AMD and OpenAI announced a six-gigawatt agreement spanning multiple generations of Instinct GPUs, with the first one-gigawatt deployment of MI450 parts starting in the second half of 2026. AMD issued OpenAI a warrant for up to 160 million shares — roughly a tenth of the company — vesting on deployment milestones, on AMD share-price targets, and on OpenAI hitting the technical and commercial milestones needed to enable deployments at scale. Reported terms put the strike at one cent per share with a $600 ceiling, which is why the announcement added about $80 billion to AMD's market value.

Neither company would put a dollar figure on the deal, calling it worth billions. The timing is the sharper detail: it landed hours after OpenAI's DevDay keynote, extending a run in which OpenAI has committed to roughly a trillion dollars of compute build-out in a matter of weeks.

WHY IT MATTERS

For anyone buying inference, this is the first credible second source of frontier-scale accelerator capacity from late 2026, which makes MI450 a planning assumption rather than a hedge — and makes ROCm and vLLM portability work worth doing on your own schedule instead of under pressure. It also reprices compute commitments as balance-sheet instruments, since a supplier is taking equity for volume; expect that structure to shape multi-year availability, allocation and discounting across the market.

OpenAI: AMD and OpenAI announce strategic partnershipAMD 8-K exhibit: 6 GW agreement and warrant terms
MON · Oct 6, 2025copyrightopenaisoraregulation

MPA to OpenAI: stopping Sora 2 infringement is your job, not the rightsholders'

On Monday the Motion Picture Association urged OpenAI to take immediate and decisive action against Sora 2 outputs that it says infringe its members' films, shows and characters. MPA CEO Charles Rivkin argued that OpenAI must acknowledge it remains their responsibility — not rightsholders' — to prevent infringement on the Sora 2 service, and that well-established copyright law applies. The statement followed Sam Altman's Friday blog post reversing Sora's copyright posture: rather than making studios opt out, OpenAI said it will give rightsholders more granular control over their characters and share video revenue with those who allow generation.

The collision was scheduled by OpenAI itself. Sora 2 launched on 30 September, was the top free app on the App Store within days, and on Monday morning — hours before the MPA's statement — OpenAI used DevDay to open Sora 2 to developers via the API. The failure mode was already public: clips featuring Mario, James Bond and other protected characters, which is what turned a product launch into an industry-liability question.

WHY IT MATTERS

This is the liability template for any product that generates from user-supplied IP. Opt-out pushes the enforcement cost onto rightsholders and reliably draws organized industry pushback; opt-in plus revenue sharing converts the same content into a licensed input with a metered price. If you ship generation features, the decision-relevant work is provenance tracking and per-character policy controls — we'll iterate is not a defense once a rights-holder trade association is the counterparty.

CNBC: OpenAI's Sora 2 must stop allowing copyright infringement, MPA saysBusiness Insider: Altman walks back the Sora copyright opt-out policy