NPDev Daily AI Briefing - October 29, 2026
Microsoft cut its 2026 AI capex forecast to about $170B and Alphabet held guidance while posting negative free cash flow, both citing compute that arrived later than planned. Google shipped Gemini 4 Carbon to cyber defenders only, OpenAI said at its Berlin DevDay that it has no video-input model and shut Sora because customers wanted coding and agents, and Anthropic put spend-attribution reports in front of the CFOs paying for tokens.
Written by NPDev Labs with AI assistance and reviewed before publication. How we work.
Microsoft trims 2026 AI capex to $170B as Alphabet posts negative free cash flow
Microsoft's fiscal Q1 (quarter ended Sept. 30) came in at $92.3B of revenue, up 19%, with Azure at $33.2B, up 48% — the first time Microsoft disclosed Azure revenue in dollars. Capital expenditure including finance leases was $45.2B. CFO Amy Hood cut calendar-2026 capex to about $170B from $175B and said some third-party GPU capacity slipped to the second half of the fiscal year, while guiding Azure growth of 49–50% at constant currency: "We still have demand that exceeds supply." Alphabet held capex guidance at $195–205B but missed on revenue and earnings, reported more than $9B of negative free cash flow, and fell about 10% after hours; its CFO said outside compute "came online only at the very end of Q3." CoreWeave, Nebius, IREN and Oracle all fell.
Why it matters: The constraint has moved from demand to delivered capacity: compute you assumed would be rented in Q4 is now a 2027 problem, and GPU-as-a-service middlemen are a measurable dependency risk. Treat reserved capacity as a contract with delivery dates and penalties, keep a second provider warm, and model inference cost at the higher of spot and reserved pricing rather than the launch price.
Status: Confirmed first-party earnings releases, Oct. 28. Capex guidance and the delay explanation are CFO statements on the earnings call; share moves and comparisons are reported by CNBC.
Sources:
-
-
-
OpenAI says it has no video-input model and no Sora because customers want coding and agents
At DevDay Exchange in Berlin on Oct. 28, OpenAI's developer lead Romain Huet said the Dots assistant should reach Europe more widely "in a few weeks" after talks with regulators; personal Pro accounts in the EEA, Switzerland and the UK are still locked out, though some business accounts have access. OpenAI said weekly active Codex users in Germany are up twentyfold since January and that nearly one in five desktop users now use skills for documents, presentations or spreadsheets. Asked why there is still no model that understands video sequences the way Gemini does, management said customers barely ask for it, so it has not gone beyond single-frame analysis. It again attributed Sora's March shutdown to extreme resource demands and said customers wanted coding models and agents, without ruling out a return.
Why it matters: This is unusually explicit roadmap reasoning: capability investment is following revenue concentration in coding and agents, not multimodal breadth. If your product needs sequence-level video understanding, Gemini is currently the only frontier option, and anything built on Sora has a demonstrated shutdown precedent — keep a portability plan for every OpenAI-only surface.
Status: Reported from on-stage remarks and the Q&A by The Decoder, Oct. 28; OpenAI has not published a post covering these answers. The Europe timing is a statement of intent, not a date.
Sources:
-
Anthropic ships spend-attribution reports as token costs land in engineering budgets
Anthropic announced reporting tools that let managers see a team's work over a date range, what it produced, what it cost and where workflows broke down, with recommendations and pooled group budgets; the data is not used for model training by default. The launch follows its S-1, which showed an operating profit in Q2 and 900% net revenue retention. A Jellyfish study of agentic coding at 276,000 developers found Anthropic-based tools behind nearly a third of proposed code changes, with spend up 10x since January — now roughly 2% of total engineering cost.
Why it matters: AI coding spend is big enough to need cost-per-outcome accounting rather than seat counts. Instrument cost per merged PR and per completed task now, so the next budget conversation is about routing and context hygiene instead of a 10x line item nobody forecast.
Status: Product features confirmed first-party; S-1 financials and Jellyfish figures as reported by TechCrunch, Oct. 28. Jellyfish data is a vendor survey of its own customer base.
Sources:
-
-
Gemini 4 Carbon tops a third-party index, but ships first to cyber defenders
Google announced Gemini 4 Carbon on Oct. 28, two weeks after Gemini 4 Argon. It ranks first on Artificial Analysis Intelligence Index v5.0 at 59 and claims 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, plus a jump of 89 Elo to fourth on WebDev Arena. Starting this week it is available to a subset of cyber defenders, phasing wider through the Fairwind program, while Google engages with the U.S. government's pre-release testing process. Developers, businesses and consumers get it "as soon as possible." Introductory pricing is $2/$10 per million tokens (95% off cached input), rising to $4/$20.
Why it matters: The strongest model of the week is one you cannot call. Cyber capability is being staged to defenders first, which shortens the window between a model finding a bug and attackers having something similar — expect patch cadence pressure. Benchmarks are Google's own except the Artificial Analysis index; treat the gap between "announced" and "available in your region and tier" as the planning number.
Status: Confirmed first-party announcement. Limited access, not general availability; no developer or consumer date published. Pricing is the published introductory rate.
Sources:
-
-
Meta's Muse personal agent opens in Singapore with a local connector catalog
Meta made Muse available in Singapore on iOS, Android and muse.ai, with connectors to Gmail, Google Calendar and Instagram plus local partners including Shopee, Grab, Singapore Airlines, Agoda, ShopBack, Klook, Trip.com and Lazada; some connector features roll out progressively and AI glasses support is coming. Muse runs on a "Muse Secure VM" that stores data and credentials for connected services outside the agent's view, and asks before sensitive actions such as sending mail or buying. It is free for most uses with subscription tiers.
Why it matters: Consumer agents are shipping market by market with a connector catalog, not as one global model launch — which is also how you should roll out an internal agent. The secure-VM pattern (credentials held outside model context, explicit confirmation per sensitive action, per-connector scope) is the reference architecture to copy before wiring an agent to production systems.
Status: Confirmed first-party, dated Oct. 27–28 on Meta's newsroom. Availability is Singapore only; the connector list includes partners not yet fully live.
Sources:
-
Kimi K3.1 Pro weights are still pending, leaving DeepSeek's MIT snapshot as the open default
Moonshot's Kimi K3.1 Pro, launched Oct. 23 with a 1M-token context and 65.7% on Terminal-Bench 4.0, is API-only for now: the company says full weights land by Nov. 4, and as of Oct. 29 its Hugging Face organization still lists Kimi-K3 (Sept. 2) as the newest repository. DeepSeek's V4.1-Flash-1017 remains downloadable under an MIT model card at $0.30/$1.20 per million tokens at peak, half that off-peak. Mistral Large 4 is still a public preview whose weights link reads "coming soon" two days before its end-of-October deadline; GLM 5.3 and GLM 5.3 Flash are generally available on Bedrock and Mistral.
Why it matters: "Open" is a property of a specific checkpoint, not a family. If a compliance or air-gap requirement depends on downloadable weights, check the repository and license file yourself on the day you need it; an announced date and an API endpoint are not the same thing. Today's realistic self-hosted frontier option is still DeepSeek's MIT snapshot, not Kimi.
Status: Vendor pages and the Hugging Face organization listing checked Oct. 29; launch dates and pricing cross-checked with a secondary tracker. Weight-release dates are vendor promises, not deliveries.
Sources:
-
-
Google opens a Connected Apps Platform so one MCP server reaches Search and Gemini
Google launched a preview of its Connected Apps Platform: developers integrate a cloud-hosted MCP server or Android MCP through AppFunctions, then register, test and submit the app in the Google Connected Console, choosing whether it goes live in Search, the Gemini app or both, with traffic, latency and error tracking afterwards. More than 50 connected apps are live, with Adobe, Experian, Localiza and Peloton in Gemini and Calm coming to both surfaces. Access is phased, starting with U.S. developers on a waitlist.
Why it matters: MCP is turning into the distribution contract for assistant surfaces: one server, two consumer entry points. If you already run an MCP server internally, publishing it is now a product decision with review queues, per-surface behavior and its own observability requirements — and no broad-availability date.
Status: Confirmed first-party, Oct. 28. Public preview, U.S. developers only; Google gave no general availability timeline.
Sources:
-
-
Deepgram buys Daily, the company behind open-source Pipecat
Deepgram said on Oct. 29 that it is acquiring Daily, which maintains Pipecat, the open-source framework that orchestrates voice and video streams across speech-to-text, LLM and text-to-speech stages to cut latency. Terms were not disclosed. Both CEOs said Pipecat stays open source and vendor-neutral. Daily had raised more than $62M; Deepgram raised a $130M Series C and reached unicorn status earlier this year. Both companies came out of the same Y Combinator batch in 2016.
Why it matters: The orchestration layer for voice agents is consolidating under a model vendor, which is exactly where API pricing and roadmap leverage gets tested. A public "open source forever" commitment is not a license change, so pin versions, keep STT/TTS providers swappable behind your own interface, and watch the Pipecat governance and release cadence over the next two quarters.
Status: Confirmed by both companies to TechCrunch, Oct. 29. Deal terms undisclosed; the open-source commitment is a public statement, not a contractual guarantee.
Sources:
-
Bottom line
Capex guidance stopped rising and capacity arrived late, so for the next two quarters the binding constraint is delivered compute rather than demand. Frontier capability is arriving gated — Carbon to cyber defenders, Muse by country, Kimi by promised weight date — while the agent platform layer consolidates around MCP and voice orchestration. For engineers the actions are unglamorous: attribute token spend to outcomes, verify open-weight claims at the repository, and keep every vendor behind a swappable interface.
As of: October 29, 2026, 15:05 UTC