AI Daily — Apr 27, 2026

Two of the AI industry's founding contracts came undone before lunch. Microsoft and OpenAI ended the exclusivity that defined the last four years of enterprise AI, and China's National Development and Reform Commission ordered the unwinding of Meta's $2 billion Manus deal with a one-paragraph statement. The models keep getting better, but the scaffolding around them — who owns whom, who can buy whom, whose cloud they run on — is being renegotiated faster than product cycles can keep up.

• DeepSeek V4 Pro & Flash (DeepSeek): MIT-licensed, 1M-token context, runnable on Huawei Ascend chips; V4 Flash priced at $0.14/$0.28 per million input/output tokens.
• Hermes Agent v0.11.0 (Nous Research): Rewritten React TUI, dashboard plugin, expanded inference providers, and same-day support for both DeepSeek V4 and GPT-5.5.
• Chrome Prompt API (Google): On-device Gemini Nano calls directly from the browser — no server round-trip, no data exfiltration after model download.
• ML Intern (Hugging Face): Open-source CLI "AI intern" for ML work — researches papers, runs experiments, iterates up to 300 steps. (See research links; HF release coverage.)

Microsoft's license to OpenAI intellectual property is no longer exclusive. OpenAI continues to pay Microsoft a revenue share through 2030, but Microsoft stops paying one back, and the cap is now "independent of OpenAI's technology progress" — the contractual AGI trigger clause that would have suspended payments if a panel of independent experts declared OpenAI had achieved artificial general intelligence was removed.

The unlock matters more than the unwinding. OpenAI can now serve "all of its products" to customers across any provider, including Amazon and Google. Amazon said AWS will offer OpenAI models through Bedrock within weeks, with details expected at a San Francisco event Tuesday. D.A. Davidson analyst Gil Luria told Bloomberg that AWS and Google Cloud customers "have been limited in their ability to integrate OpenAI's products because of the exclusive relationship and will now be more likely to consider OpenAI alongside Anthropic."

If this works, OpenAI becomes a true multi-cloud vendor and Microsoft retains a guaranteed annuity through 2030 without the exclusivity tax that was hurting Azure's competitive standing. Watch Microsoft's earnings on Wednesday — Azure grew 38% year-over-year in the most recent quarter; Microsoft had previously cited capacity constraints around 40%. Anything below 35% guidance would intensify investor scrutiny of Azure's competitive moat.

China's National Development and Reform Commission ordered all parties to withdraw from Meta's acquisition of Manus, a Singapore-registered AI agent startup with Chinese founders. The commission did not elaborate. Manus's website still says it "is now part of Meta" — indicating Beijing is unwinding a transaction both parties considered closed.

The backstory is harsher. The Financial Times reported in March that Manus CEO Xiao Hong and chief scientist Ji Yichao were summoned to Beijing and barred from leaving the country during the review. The South China Morning Post reports the case was screened against export-control and outbound-investment rules — treating agentic AI architectures as a controlled technology, similar to advanced chips.

Lian Jye Su, chief analyst at Omdia, told NPR that China is "willing to play hardball when it comes to AI talents and capabilities." If Meta walks away, every Western firm eyeing a Chinese-founded startup — even one re-domiciled to Singapore or Dubai — will assume higher geopolitical M&A risk. If Meta sues or restructures, watch for Beijing to escalate before the Trump–Xi summit in the coming weeks. The observable signal: whether other Chinese-founded startups in neutral hubs pause fundraising or quietly restructure leadership by the end of May.

OpenAI confirmed Sora's web and app experiences were discontinued April 26, with the API scheduled to be discontinued on September 24, 2026. The official reason: "strategic shifts and complex copyright challenges" — corporate language for a risk-reward calculation that no longer pencils out.

Sora was technically impressive: physical-interaction modeling, character consistency across angles, the kind of output that distinguished it from earlier video generators. None of that mattered against the legal exposure of training data nobody wanted to defend in discovery. The lesson the industry is internalizing: build with licensed data or don't build at all. Expect the next wave of video tools to ship with provenance disclosures attached, and expect standalone generative-media apps to continue collapsing into enterprise content pipelines where the licensing chain is auditable.

Three weeks after Mercor disclosed a breach, the scope keeps expanding. The threat actor, TeamPCP, compromised the LiteLLM open-source library's PyPI credentials in late March and injected a three-stage backdoor that ran on installation. From there, attackers exfiltrated roughly 4TB from Mercor — the AI data-training startup last valued at $10 billion.

The contents are the story. Per Oravys, the dump includes voice samples for over 40,000 contractors, paired with government-issued ID documents and high-resolution interview video. The Wall Street Journal reported in February that off-the-shelf voice cloning needs about fifteen seconds of clean audio; Mercor recordings average two to five minutes of studio-quality speech per contractor. That's a ready-made fraud toolkit with the credentials attached.

Per Halborn's analysis, the leak reportedly includes records of how OpenAI selects training data, how Meta builds reinforcement learning, and how Anthropic implements safety labels — methodological IP that differentiates each frontier lab. A source told Staffing Industry Analysts that Meta has paused its work with Mercor. Watch for the FTC or EU regulators to issue contractor-security guidance within thirty days; if they do, the cost of building with licensed human-generated training data will rise across the board.

DeepSeek's technical report, dissected this weekend across Latent Space, Artificial Analysis, and arena.ai leaderboards, makes V4 Pro the #2 open-weights reasoning model at an Artificial Analysis index of 52, behind Kimi K2.6. V4 Pro pricing is $1.74/$3.48 per million input/output tokens; V4 Flash is $0.14/$0.28. Per arena.ai, V4 Pro ranks roughly 30 places higher than Flash but costs 12× more.

Architectural disclosures matter as much as benchmarks. Per the technical report, the model uses Compressed Sparse Attention and Heavily Compressed Attention to reduce KV cache memory by 8.7× versus V3.2 at 1M tokens — meaning 9.62 GiB per sequence instead of 83.9 GiB. NVIDIA reported V4 Pro hits over 150 tokens per second per user on Blackwell GB200 NVL72 in early tests. DeepSeek says the model also runs on Huawei Ascend with 1.5–1.73× acceleration ratios, and pricing could fall further once Ascend 950 supernodes scale in H2 2026.

The political read: this is the first credible open-weights model that bridges NVIDIA and Huawei stacks at production scale. If Chinese enterprises and developers across emerging markets standardize on V4-class economics, premium API pricing could collapse for everything except the highest-margin frontier workloads.

Google's Chrome Prompt API hit Hacker News this morning — not because it's new (the docs were last updated in September 2025) but because developers are belatedly noticing what it implies. The API runs Gemini Nano locally, requires at least 22 GB free storage and either 4+ GB VRAM or substantial CPU, and sends no data to Google or third parties after the model downloads.

The signal isn't the API; it's the timing of the attention. On-device inference is becoming a real distribution channel for lightweight agents, summarizers, and privacy-sensitive workflows that don't justify a server round-trip. If Apple ships an equivalent in the next Safari release — or if developers start shipping production features against Chrome's API — the browser could become the most-installed AI runtime on Earth, and a meaningful slice of agent traffic would stop touching the cloud entirely.

Two founders held in Beijing while their Singapore acquirer waited by the phone; an AGI clause deleted by lawyers before any model could trigger it; forty thousand voice samples and matching IDs auctioned to whoever wanted to be those people on the next bank call. The contracts that built this industry are getting unwound faster than the models that justified them — and the ones writing the new ones aren't all in San Francisco. Onward.

Forward this to whoever's still pretending the AGI clause meant something.