AI news, 22 October 2026: the build-out showed up on the balance sheet.

Meta spent $34.91 billion on capital expenditure in a single quarter, reported free cash flow of negative $2.12 billion, and in the same release announced a new class of non-voting shares it can sell to fund what comes next. Google Cloud started selling OpenAI's models and let enterprise customers pay for them out of commitments they had already made to Google. Salesforce agreed to pay about $5.66 billion for Braze. Three of today's five stories are about who finances AI and how, and none of them is a model launch. The other two are about what you can actually download: DeepSeek published the inference engine behind its production service, and NVIDIA open-sourced most of a robotics stack built around a humanoid that does not ship until 2027.

Every date and figure below comes from the primary source linked in its section, except where a section says otherwise. AcquihireTech writes this roundup for Indian service businesses doing ₹10L+ a month, so each item is filtered through one question: does it change what you do this month? The window is 20 to 21 October. The note near the bottom lists what was cut and why.

1. Google Cloud started selling OpenAI's models, billed against Google's own commitments

What happened. On 21 October Google Cloud published an announcement by Michael Gerstenhaber, VP of product management for Gemini, putting three OpenAI models into its model garden. They are GPT-6 Astra, described as designed for “demanding work across coding, computer use, professional tasks, and scientific research”; GPT-6.1 Sol, “near-Astra intelligence at lower cost”; and GPT-6 Luna, “a fast, cost-efficient model for everyday work at scale”. All three are “rolling out now in private preview to select enterprise customers”, with access by interest form. From the same day customers can build with them through Google's APIs and connect local Codex CLI and Desktop workflows to Google's Gemini platform for inference. The billing arrangement is the part Google spells out: because the models are transacted through Google Cloud Marketplace, “all usage appears on a single Google Cloud invoice, draws down existing cloud commitments, and counts toward Flexible Savings Plans (FSP) across model providers”. Availability in Gemini for business and Google Antigravity is “coming soon”. Google claims it is “now the only hyperscale cloud provider offering access to all three leading closed source LLMs” alongside more than 200 first-party, third-party and open models; that is Google's own framing and the post names neither the other two providers nor a general-availability date, and publishes no pricing.

Why it matters. The news is the invoice, not the model list. A large enterprise signs multi-year committed spend with a cloud provider, and commitment it fails to consume is money already spent. A sentence that lets OpenAI usage draw that balance down turns an awkward procurement conversation into a line item somebody has already approved, which is a far shorter path into a company than a better benchmark score. Be exact about what shipped, though. Private preview, select customers, an interest form and no published price is not a product you can buy, and Google says so plainly in its own post. The claim about being the only hyperscaler with all three closed-source leaders is also a category the company drew around itself.

What it means for a service business. You almost certainly do not hold a committed-spend contract with Google Cloud. Your larger clients might, and that changes a conversation you have probably had. When a client's IT team tells you which model you are allowed to use, the reason is now at least as likely to be where the spend lands as what the model can do, and it is worth asking which it is before you design around whichever API you prefer. For your own stack this month, nothing: an unpriced private preview is not something to build on, and the cost of waiting is zero.

2. Meta's free cash flow went negative, and it created a share class to fund the next round

What happened. Meta reported third-quarter results on 21 October. Revenue was “$64.60 billion, an increase of 26% year-over-year on both a reported and constant currency basis”. Capital expenditure, including principal payments on finance leases, was $34.91 billion for the quarter, and the reconciliation in the same filed release puts free cash flow at negative $2.12 billion. Full-year 2026 capital expenditure guidance is now “in the range of $135-145 billion, narrowed from our prior outlook of $130-145 billion”: the floor moved up by $5 billion and the ceiling did not move. There is no 2027 number in the release, which says only that Meta expects to keep investing aggressively and is still developing that budget. Separately, the board approved a publicly listed, non-voting Class C common stock with an effective date of 19 January 2027. The information statement sets out the mechanics: a dividend of one Class C share for each outstanding Class A share and the same ratio for Class B, after which Meta expects “the market price for the shares of our Class A common stock will reflect the effect of a two-for-one stock split”. The stated purpose is a non-voting security the company “may use for capital raising”, plus employee compensation, potential acquisitions and other general corporate purposes. Control is untouched: as of 1 October Mark Zuckerberg held approximately 60.6% of total outstanding voting power on approximately 13.4% of economic interests, and under a Founder Alignment Agreement 20% of the Class C shares issued to his holder group are freely saleable while the other 80% are “Stapled Class C Shares”, sellable only alongside a corresponding sale or conversion of Class A or Class B shares, or with Stapling Credits.

Why it matters. Guidance ranges are forecasts and can be revised. Negative free cash flow is an accounting fact that has already happened, and $34.91 billion of quarterly capital expenditure against $64.60 billion of quarterly revenue is the plainest published figure anyone has for what frontier AI infrastructure costs the company building it. The share class is the more informative half of the day. A company that raises its capex floor and in the same release creates a non-voting instrument explicitly available for capital raising is signalling where it expects the next phase of funding to come from, and doing it in a structure that dilutes economics without touching votes. One thing not to overstate: Meta has said it may use the new class for capital raising, not that it will. The filing creates the option, and the option is the news.

What it means for a service business. Your inference bill is a retail slice of this spending, and the condition on display here is a vendor investing well ahead of the cash its operations generate. That is historically when introductory pricing stops being introductory. So if a model API is load-bearing in how you deliver work, find out this month what the same task costs on your second choice, and write the number down. Discovering it during a price change is the expensive way. Second, there is a cheaper way to follow this than reading opinion pieces about whether AI is a bubble: capital expenditure against operating cash flow, in each vendor's own quarterly release. It is published, it is quarterly, and it is not anybody's opinion.

3. Salesforce agreed to pay about $5.66 billion for Braze

What happened. On 20 October Salesforce and Braze announced a definitive agreement for Salesforce to acquire Braze for “approximately $5.66 billion in enterprise value”, with Braze Class A holders receiving “$47.50 in cash per share”, according to the announcement Braze filed with the SEC. The transaction “is expected to close in Salesforce's fiscal year 2028”, subject to Braze shareholder approval and regulatory approvals. Salesforce's rationale, in its own words: Braze's “real-time engagement and AI decisioning capabilities are expected to complement all four layers” of its architecture, combining them with “Salesforce's customer context and Agentforce would help businesses” make engagement more relevant, and “Braze is expected to strengthen Marketing Cloud's cross-channel personalization” with real-time optimisation across channels.

Why it matters. Read the close date before the price. Salesforce's financial year ends on 31 January, so fiscal 2028 does not begin until February 2027, which puts this deal at least a year out and makes today's announcement the only part of it that exists. In the meantime two separate companies keep selling you two separate products. The price is worth a second look too, because messaging orchestration and decisioning is not a capability Salesforce lacks; it sells one. Paying $5.66 billion for somebody else's says the engagement layer underneath Agentforce needed to be better than what was there, sooner than it could be built.

What it means for a service business. If you or your clients run Braze, nothing breaks for at least a year and existing contracts run to close. But a tool in the middle of an acquisition is a poor moment to sign a three-year renewal, and the fix is one clause: shorten the term, or get change-of-control language in writing. If you run Marketing Cloud, the roadmap you were sold now has a dependency that did not exist last week, and the integration plan will not be published for months. The wider lesson is the one that keeps repeating in this category: the messaging tool on top of your customer data gets acquired, and the data underneath it does not have to move when that happens. AcquihireTech builds enquiry pipelines that way deliberately, with the record of who enquired and what happened next sitting in a store the client owns, so changing the tool that sends the message is a swap rather than a migration.

4. DeepSeek open-sourced its inference engine, minus the parts that make it a service

What happened. DeepSeek published DeepInfer under the Apache 2.0 licence, reported on 20 October. Its README calls it “the high-performance inference framework that powers DeepSeek's large-scale production services”, and says this open-source release provides offline inference. It supports one model, DeepSeek-V4.1-Flash, with DeepSeek-V4.1-Flash-1017 as the default checkpoint, on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. Prerequisites are Linux, GCC 13 or newer, Python 3.12, and either CUDA 13 or CANN 9.2. What is in the box: optimised kernels, attention data parallelism, balanced context parallelism, mixture-of-experts expert parallelism with EPLB, DSpark speculative decoding, SWA bounded replay, an asynchronous zero-overhead batch scheduler enabled by default, batch-invariant and deterministic sampling, and structured outputs. What is not in the box, in the README's own words, is that a complete inference service requires components “not included in this release”: a KV cache system, prefill and decode disaggregation, and load balancing across data-parallel replicas, though it adds that most of the engine functionality needed for disaggregation is already implemented. Performance appears only as prefill and decode Pareto charts, measured on 4 SM100 GPUs for prefill and 64 for decode, with no numbers in the text. The README carries no release date; the 20 October date comes from dated reporting rather than from DeepSeek.

Why it matters. The exclusions are the story. Those three missing pieces are the hard parts of serving a model to real traffic at a sane cost, and keeping them back means the release is genuinely useful to study and to run offline while remaining a long way from a drop-in replacement for anybody's hosted API. That is a deliberate line, and where a company draws it tells you more about its business than the licence does. One other detail deserves a second look: a single release that treats Huawei's Ascend 950 as a first-class target alongside NVIDIA silicon, published in the open, is a supply-chain fact as much as a software one.

What it means for a service business. Almost nobody reading this should be running their own inference, and this README is the clearest available argument for why: the part that was given away is the easier half. So when a vendor pitches self-hosted open models as a saving, two questions settle it. Who operates the KV cache and the load balancer, and what is the monthly cost of the person who does? Get both answers as numbers. One feature here is worth raising with your existing hosted vendor, though. Batch-invariant and deterministic sampling is the difference between a quote generator that returns the same answer to the same input twice and one that quietly does not, and if you have ever had a client query two different numbers for the same job, that is the setting you were missing.

5. NVIDIA put a reference humanoid on stage and open-sourced the simulation around it

What happened. At GTC Berlin on 21 October NVIDIA announced a set of Isaac platform releases. The headline hardware is Agile Robots' Agile ONE SN, a full-stack reference humanoid built on NVIDIA Isaac GR00T and powered by Jetson Thor: 1.40 m tall, 40 kg, 31 degrees of freedom, 3 kg payload per arm, dual WUJI Hand 2 five-finger hands with 20 degrees of freedom each, 1 kHz force control and tactile sensing, plus head and wrist cameras, lidar and inertial measurement units. The Jetson Thor compute on board is rated at 2,070 FP4 teraflops with 128 GB of unified memory. Integration with NVIDIA's Halos safety stack is described as planned. The robot is “scheduled to be available in early 2027”. The software is the part available now: Isaac Lab 3.0, an open-source, agent-ready framework for training and refining robot policies with multi-GPU workflows on the Newton and PhysX physics engines, is “now generally available”; Isaac OS is an open-source runtime built on NVIDIA Holoscan that connects AI, sensors, controls and robot hardware; and the SimReady Central Library puts more than 2,000 simulation-ready robotics and factory assets on Hugging Face, alongside the Human 1M Dataset, whose full version goes to Cosmos Coalition partners. NVIDIA's claim that the platform cuts setup time “from weeks to days” is the vendor's own and is not independently tested.

Why it matters. The humanoid is the photograph; the simulation stack is the product. Sort the announcement by what carries a date and the shape is obvious: Isaac Lab 3.0 is generally available today, the robot arrives in early 2027, and the safety integration has no date at all. This is the same sequence NVIDIA ran in AI, which is to give away the framework, let everyone's workflow standardise on it, and sell the accelerator underneath. It is a reasonable bet that it works twice.

What it means for a service business. Nothing to act on this quarter, and that is worth stating rather than manufacturing relevance. There is one real read-across. If your clients are in manufacturing, warehousing or logistics and somebody pitches them humanoid automation in the next eighteen months, two facts from this release are a complete answer to an aggressive delivery date: the reference platform is not purchasable until early 2027, and its safety-system integration is unscheduled. Ask any vendor promising sooner which hardware they are shipping and which safety certification it holds.

The window, and the stories we cut

Today's window is 20 to 21 October, and all five items fall inside it. Four were verified at a primary source opened during this run: Google Cloud's own blog post, Meta's filed earnings release together with its information statement on EDGAR, the joint Salesforce and Braze announcement as filed by Braze with the SEC, and NVIDIA's newsroom release. The DeepInfer section is drawn from DeepSeek's own README, which is primary for every technical claim in it; only the 20 October release date comes from dated reporting, because the README carries none.

Five stories were cut. OpenAI's acquisition of Orchestra Bio, announced on 20 October and reported by several outlets that link to OpenAI's own post, is absent because openai.com returns HTTP 403 to every automated fetch from this pipeline, as it did on previous runs, so the announcement could not be read at its source. A research preview circulating as “Claude Duet” is absent because it does not appear on Anthropic's news index, which was checked during this run. A claim that a draft Anthropic S-1 shows $16.3 billion of first-half 2026 revenue is absent for the same reason: no filing or company statement supporting it could be found. Mistral's Timestral forecasting models and its statement of intent with CERN are both real and both dated 15 October, which puts them outside this window rather than in doubt. And a 140-page proof of the Kannan–Lovász–Simonovits conjecture attributed to Microsoft Research appeared in a single aggregator summary with no reachable paper or institutional post behind it.

The short version of this edition: find out what your most-used AI task costs on your second-choice model, before you need to know. If a martech tool in your stack is mid-acquisition, shorten the term or get change-of-control language written in. Keep the record of who enquired in a store you own, so the messaging tool on top of it stays replaceable. And if anyone quotes your client humanoid automation for 2026, the reference platform ships in early 2027 and its safety layer has no date. If that last exercise turns up a vendor nobody in your team can price a replacement for, that is roughly where an AcquihireTech Systems Audit starts.