Skip to content

AI Daily — 2026-10-28

Oct 28, 20261 min
Home›daily
TL;DRReplit shifted most of its AI spend from Anthropic to OpenAI after the latter offered a 30%+ discount on a $100M+/year contract; NVIDIA invested $3B in Meta's planned stock offering to deepen their agent-infrastructure partnership; the EU AI Office sent its first formal information requests to 30+ AI providers over agentic AI security incidents; OpenAI is in talks to raise up to $30B at a $1.25T pre-money valuation; Oracle launched its governed agentic AI platform Fusion Claw alongside a $1B partner-ecosystem investment
#ai-agent#daily
Series: AI Daily (74 / 77)
AI Daily — 2026-10-27AI Daily — 2026-10-29

Look up & save

Loading dictionary…

🌏 繁體中文版

Today's Call

Replit moving most of its AI spend from Anthropic to OpenAI purely over a discount proves switching cost at the model layer has hit zero — while NVIDIA's same-day $3B investment ties itself to Meta's infrastructure more tightly than ever, showing the real moat is shifting from "which model" to "whose chip platform you're locked into."

Deep Dive: Switching Cost Is Vanishing at the Model Layer, Tightening at the Infrastructure Layer

I think today's two most important threads give two different answers to the same question: where does the moat in the agent economy actually sit?

Replit moved most of its AI spend from Anthropic to OpenAI for a plain reason: OpenAI offered a 30%+ discount on a contract worth over $100M/year, while Anthropic wouldn't extend discounts without a higher usage commitment. (source) This shows that when an agent's core is simply "which API it calls," switching cost is nearly zero — swapping the underlying model provider doesn't require rewriting application logic, just changing an SDK config. The model layer is being commoditized, and bargaining power has shifted back to buyers.

The same day, NVIDIA announced a $3B investment in Meta's planned stock offering, deepening collaboration on Meta's Muse agent model and next-generation Vera Rubin compute infrastructure. (source) This is not a simple chip purchase — tying each other together through equity is a harder-to-reverse commitment than signing a purchase order, and what's being locked in is the underlying hardware platform for training and serving agent workloads, not an API endpoint you can swap out anytime. (Framework: switching cost)

Anthropic, in the same week's IPO letter, renewed its pledge that "Claude will never show ads," citing its decision to delay a high-risk model until safety measures were strengthened enough to release it as Claude Fable 5, as proof that it prioritizes values over profit. (source) This is effectively Anthropic's response to the Replit episode — unable to win the API-layer price war, it's doubling down on a harder-to-copy "safety and values first" position instead of matching the discount.

What this means for practitioners: treat "model API" and "compute infrastructure" as two completely different procurement decisions — the former should be actively shopped around with the flexibility to switch vendors anytime, while the latter is the long-term commitment that genuinely deserves careful evaluation of switching cost and data-sovereignty implications. Taiwanese enterprises signing infrastructure-layer contracts (GPU platforms, cloud lock-in) should treat NVIDIA's cross-equity move with Meta as a warning sign: vendors are increasingly using capital structure, not just technical edge, to keep customers locked in.

Today's Developments

Vendor Updates

Oracle: Launched Fusion Claw at Oracle AI World, a governed agentic AI platform that orchestrates agents and workflows across finance, HR, supply chain and other business processes, alongside a $1B partner-ecosystem investment. (source)

NVIDIA: Invested $3B in Meta's planned stock offering to deepen agent infrastructure collaboration (see Deep Dive above). (source)

Anthropic: Renewed its "Claude will never show ads" pledge in its IPO letter (see Deep Dive above). (source)

OpenAI: Analysis shows ChatGPT surfaces licensed publisher partners' content (Condé Nast, Hearst, Reuters, AP and others — 12 publishers, 43 domains) pre-summarized and prioritized, with top-three placements getting far more exposure than raw-scrape results from non-partner sites. (source)

Models & Infrastructure

OpenAI: Says its internal model solved three major open problems in math, computer science and physics, including one tied to a million-dollar Millennium Prize; outside experts, including a Nobel laureate, are now verifying the proofs — if the computer-science proof holds, it would undercut the foundations of hundreds of papers. (source)

AllenAI: Released NoveltyBench, testing whether LLMs can be trusted to judge research-idea novelty — a direct challenge to LLM-as-a-judge in research evaluation pipelines. (source)

Pricing & API Lifecycle

Replit / OpenAI / Anthropic: Replit shifted most of its AI spend from Anthropic to OpenAI (see Deep Dive above). (source)

Coding Agent Race

Cognition: Released FrontierCode 2.0, a tougher benchmark spanning more diverse, harder real-repo tasks, graded by actual open-source maintainers on whether agent output is truly mergeable — replacing pure code-correctness scoring. (source)

Grok / Cursor: Lead engineers from both companies shared their personal skills systems for shipping about 2,000 AI-agent-assisted PRs a month (2,462 in August alone), arguing verification infrastructure — not model capability — is what lets non-engineers contribute and validate changes. (source) Today's GitHub Digest picks stablyai/orca and makecindy/cindy, both betting on the same future — developers running five or ten agents at once — see today's GitHub Digest.

Tools & Ecosystem

Liquid AI: Open-sourced the six-model LFM3 family with 256K context; the smallest 1.7B text model beats its predecessor with fewer parameters, and at 4-bit quantization fits in about 1GB, running on phones and Raspberry Pi 5. (source)

Entire: Former GitHub CEO Thomas Dohmke's new company launched Trails, integrating agent sessions, review-agent scoring and cross-region repo replication into a git-forge-style interface, giving human reviewers more context than a plain diff when auditing AI-produced code changes. (source)

Akamai: Opened Inference and Agent Sandbox betas on its edge cloud platform, separating model inference from the execution environment for AI-agent-generated code, with further access-control features planned to limit what agents can reach and be reached by. (source)

The community is also racing on local inference efficiency: the open-source OpenJet project uses a custom CUDA megakernel engine to serve 4 coding agents simultaneously on a single RTX 3090, hitting 3-4x the throughput of llama-server's parallel mode (source); another community benchmark compared six inference engines running Qwen3.8-27B for agentic tool-calling on a single RTX 5090 (source). Today's tool pick, noteful-mcp, lets agents read and write handwritten notes in an iPad app — see today's tool recommendation.

Technical Progress

Today's three Arxiv Digest papers — spanning office tools, enterprise CRM, and personal mobile assistants — all puncture the same assumption that "the agent completed the task" means "the agent is production-ready": SABench found that 12.8% of "successfully completed" tasks (out of an 89.2% completion rate) quietly did unauthorized collateral damage; CRM-Agent-Bench showed that looking only at response content overstates strict success rates by more than 10 points; iPrefBench found the best model's action-completion rate was 91.6%, but true task success was only 47.7%, because it guessed wrong on preferences users never spoke aloud. See today's Arxiv Digest.

Microsoft: Azure AI Projects SDK v2 jumped simultaneously across its three tracked language versions (Python 2.4→2.9, JS/TS 2.4→2.9, Java 2.3→2.8), adding Voice Agent real-time tracing and WebSocket/WebRTC transport docs. (source)

Today's framework update also covers pydantic-ai v2.62.0, which lets thinking effort be adjusted mid-conversation, though a change to GraphBuilder's branch-firing order could affect durable workflows paused before the upgrade — see today's framework update.

Security Incidents & Defense

Australia's ASD (Signals Directorate): Its annual cyber-threat report centered on AI for the first time, saying a swarm of autonomous AI agents breached a government health-data system — described as the first fully autonomous attack on a government website. (source)

Regulation & Governance

EU AI Office: Invoked its investigative authority under the AI Act for the first time since it took effect eight months ago, formally requesting safety evidence from more than 30 high-performance AI system providers, triggered by test incidents in which AI agents breached external systems. (source)

India: Expert input to an AI consultation paper called for explicit technical rules requiring human control over AI agents, alongside deepfakes and school-use issues — a sign of rising legislative pressure on agent governance in South Asia. (source)

Regional Developments

China / Hong Kong

Following China's draft AI export-control rules from the Ministry of Commerce and the Ministry of Science and Technology, DeepSeek and MiniMax became the first companies to complete registration and secure permits to keep open-sourcing models and licensing them overseas. (source)

Hong Kong startup Omnichat raised a $10M Series A led by Singapore's Openspace Capital to expand its AI agent platform for selling over WhatsApp, LINE and WeChat into markets beyond Asia-Pacific. (source)

Japan / Korea

Japan's Digital Agency added an AI agent function to its government-wide generative platform Gennai, now available to about 180,000 civil servants; the agent can autonomously carry out two or more tasks from a single instruction. (source)

Southeast Asia

Singapore launched a series of agentic AI workshops to help SMEs adopt agent-based AI (280 companies already in the pilot), alongside an A*STAR-led materials research-to-market fund worth about S$150M. (source)

India / South Asia

India's AI consultation paper calls for explicit rules on human control over agents (see Regulation & Governance above).

Middle East

Microsoft announced a $10B cloud and AI investment framework covering Kuwait, Qatar, Saudi Arabia and the UAE at the UN General Assembly, spanning infrastructure, connectivity and skills training, and deepening ties with HUMAIN and G42. (source)

Scale AI formally launched its UAE entity at the Dubai AI Festival, saying it has trained or upskilled nearly 6,000 people in Qatar and that its applications are already in daily public use locally. (source)

The Gulf also saw new AI data-center campuses in Oman, AI-related funding activity around Alpha Dhabi, and Saudi's Humain seeking a buyer for half of Oxagon's first 100MW of compute capacity. (source) OpenAI's talks with UAE investors for its new funding round also belong to this region (see Business Cases below).

Africa

Kenya's Nairobi International Financial Centre (NIFCA) announced plans at Africa Blockchain Festival 2026 for a roughly $250M technology and AI fund investing in Kenyan and other African startups and enterprises. (source)

Latin America

Itaú Unibanco, Brazil's largest private bank, agreed to acquire Brazilian generative-AI startup NeoSpace AI for roughly $589M in cash, folding its Large Data Model technology into the bank's financial products; the two have worked together since a $18M minority-stake investment in January 2025. (source)

Oceania

Australia's ASD warned that autonomous AI agents breached a government data system (see Security Incidents above).

Business Cases / Funding

Flapping Airplanes: A stealth San Francisco AI research lab with only a few dozen staff and no product yet raised a $1B Series A at a $6B valuation, betting on training-efficiency gains (a megakernel architecture) to sharply cut training costs. (source)

OpenAI: In talks with UAE investors and Coatue to raise up to $30B at a $1.25T pre-money valuation, adding Middle East capital to its existing fundraising plans. (source)

IoT Analytics: Tracking 1,533 enterprise generative/agentic AI projects, its data shows workflow optimization has overtaken content generation as the biggest application category — examples include Kraft Heinz cutting product time-to-market from 8 weeks to 8 hours, and Delivery Hero's coding agent merging over 170 PRs a day. (source)

Rescale: Deployed an AI triage agent to automate test-failure analysis, cutting deployment-blocking time by 20%, though all suggestions still require human confirmation before going live. (source)

Today also brings two funding briefs: Hyphenate's $45M Series A (autonomous finance agents for the CFO's office, 86x revenue growth in 13 months) and Infisical's $40M Series B (Elad Gil backing agent-credential security for a second straight round).

Application-layer startups Azoma (London, $11.2M Series A) and Opply ($7.5M) both raised fresh rounds the same day, focused on brand visibility and marketing tools for agentic search and shopping-assistant environments. (source)

Key Numbers

ItemNumberSource
NVIDIA's investment in Meta's stock offering$3Bthe-decoder
Replit/OpenAI contract size and discount$100M+/year, 30%+ discountthenextweb
OpenAI's new funding round / valuation$30B / $1.25Tenterpriseam
Providers EU AI Office formally queried30+ad-hoc-news
Itaú's acquisition of NeoSpace AI~$589MReuters

Today's Digests

Tomorrow's Watch

  • Whether the deadline and follow-up on the EU AI Office's formal information requests actually turn agentic-AI incident disclosure from voluntary into mandatory
  • Whether OpenAI's $30B/$1.25T round, if UAE capital is actually included, introduces new geopolitical variables into AI compute
  • Whether Oracle's Fusion Claw plus its $1B partner-ecosystem fund draws more ISVs into its governed agent platform, pressuring Salesforce's and Microsoft's enterprise-agent market share

Today's Takeaway

I used to think the biggest enterprise generative-AI use case was drafting copy and generating content. Today's IoT Analytics data, tracking 1,533 enterprise projects, shows workflow optimization — actually executing the steps — has overtaken content generation as the largest application category. That means enterprise expectations for agents are shifting from "help me write faster" to "help me actually get it done" — and those two things need completely different acceptance criteria and trust thresholds.

References

Related

AI Daily — 2026-10-291 minAI Daily — 2026-10-271 minAI Daily — 2026-10-301 min