BGAD Consulting BGAD Consulting STRATEGIES. DELIVERED.

BGAD News Flash

BGAD News Flash

Daily briefing: digital, tech and AI

23 October 2026

Anthropic has launched Claude Duet, an agent that works while you are not watching

Announced on Wednesday as a research preview, Duet runs tasks in the background on its own virtual machine rather than inside an open chat session, polls connected accounts every few hours and messages the user only when something needs attention. It opens with a short daily list of action items pulled from connected apps, runs on Opus 5.5 and ships as standalone iOS and Android apps, waitlisted to US Claude subscribers. Anthropic says a new classifier and system prompt guard against leaking sensitive data and that Duet never directly sees passwords or sign-in tokens. It lands squarely against Meta's Muse and OpenAI's Dots, and at Pro pricing of 20 dollars a month it undercuts the 100 dollar tier where Dots currently sits.

Source: The AI Daily Brief, 22 October 2026, The AI vs AI Battle for the Internet

Walmart has opened its shelves to third party AI shopping agents

Walmart launched Agent Gateway, an MCP server that lets approved outside agents search its online inventory and buy on a customer's behalf without receiving that customer's personal information or purchase history. Agent Profiles arrive in November, giving each agent its own profile inside the account so its purchases show separately and can be deleted, with either per transaction approval or a pre authorised spending limit. Agents may not exceed shopper intent, may not touch payment information, and are blocked if they fail compliance. The split with Amazon is now explicit: Amazon has been blocking agents from buying for shoppers, so the two largest US retailers have taken opposite sides on agentic commerce.

Source: The AI Daily Brief, 22 October 2026, The AI vs AI Battle for the Internet

A core PHP flaw found by AI has turned the security race into machines against machines

Two memory corruption bugs in the PHP engine, nicknamed PHP2Shell, allow pre authentication remote code execution against WordPress, Drupal and Joomla, in code that sits under roughly 70 percent of the web. The flaws were found independently twice, by a researcher at Calif using GPT-5.6 Sol under OpenAI's Patch the Planet program and by Wiz using Gemini 4 Argon and its own tooling, with about two months of coordinated remediation before disclosure. CrowdStrike chief executive George Kurtz makes the matching argument from the attacker side: agents collapse the skill barrier, so techniques that used to need elite operators are now available to unsophisticated adversaries, and some attacks now complete in seconds. The uncomfortable arithmetic is unchanged, in that the defender has to close every hole and the attacker needs one.

Source: The AI Daily Brief, 22 October 2026, The AI vs AI Battle for the Internet; No Priors, 22 October 2026, The New Apex Predator

Meta is minting a non-voting share class to pay for the AI buildout

Chief financial officer Susan Li confirmed on the Q3 earnings call that Meta plans a new class of non-voting capital stock, distributed to existing shareholders as a stock dividend and publicly traded. It is intended for capital raising, employee compensation and acquisitions while leaving the existing Class A and Class B voting structure untouched, which means Mark Zuckerberg's control is not diluted by whatever Meta raises next. Semafor reported the plan before the call confirmed it. On the same call Zuckerberg said more than 10 million people in the US and Canada have set up Muse personal agents in about six weeks.

Source: The AI Daily Brief, 22 October 2026, The AI vs AI Battle for the Internet

Meta posted a record quarter and still burned cash

Revenue reached 64.6 billion dollars, up 26 percent year on year, but capital expenditure of 34.9 billion dollars, about 54 percent of revenue, pushed free cash flow to negative 2.12 billion against operating cash flow of 32.8 billion. Reported net income of 9.86 billion and diluted EPS of 3.83 dollars were distorted by a 9.89 billion dollar legal charge under the August settlement with 52 state attorneys general over teen mental health; excluding it, EPS was around 7.40 dollars, and roughly 8.7 billion of that settlement is still unbooked. Evercore's Mark Mahaney worked through the numbers on the show. For anyone tracking whether AI capex is still comfortably funded out of operations, this is the quarter where one of the most profitable companies in the world stopped covering it.

Source: Prof G Markets, 22 October 2026, Why Elon Might Merge Tesla And SpaceX

Tesla and SpaceX have each formed a special committee to examine a merger

Musk told SpaceX investors there is a lot of merit to combining the companies, and both boards have now stood up special committees to look at it. He named the obstacle himself, that a merger would need approval from 50 percent of all Tesla shareholders, which he called quite a high hurdle, and said he personally favours it but shareholders would decide. A SpaceX IPO has been put on ice while he weighs folding Tesla, Starlink and xAI into a single public company. The context is a Tesla third quarter that beat on revenue at 28.05 billion dollars and EPS at 48 cents while deliveries fell about 2 percent and net profit dropped more than 40 percent to roughly 787 million.

Source: Prof G Markets, 22 October 2026, Why Elon Might Merge Tesla And SpaceX

CrowdStrike is building a frontier style model lab for security with Nvidia

Kurtz described a joint effort with Nvidia to train security specific models, alongside a working comparison of open weight against closed models for security work and where distillation currently stands. He also drew lessons from the recent Hugging Face incident, in which an agent exploited a vulnerability in a deployed application. His broader position is that AI, like earlier platform shifts, creates net new opportunity rather than destroying the category. Worth noting that a security vendor arguing the threat landscape has fundamentally changed is also describing its own product roadmap.

Source: No Priors, 22 October 2026, The New Apex Predator: CrowdStrike CEO George Kurtz on Adversaries Armed With AI Agents

a16z has renamed its crypto arm a16z Frontier and widened the mandate well beyond blockchain

The firm used its one thousandth episode to announce the rebrand, with Chris Dixon, Marc Andreessen and Ben Horowitz setting out a remit covering autonomous devices acting in the physical world, advances in biology and medicine, full stack companies combining software, hardware and services, networks linking people and AI agents, and blockchains and stablecoins as financial plumbing including machine to machine payments. The unit has raised close to 10 billion dollars across five funds since 2018. TechCrunch read the rename as a signal that blockchain investing has fallen out of vogue, noting the sector never extended much past financial rails and speculative assets; The Information reported that Ali Yahya, the strategy's first investor, has left the firm. This is a venture firm announcing its own book, so treat the thesis as positioning rather than analysis.

Source: The a16z Podcast, 22 October 2026, The Frontier Ideas a16z Wants to Fund

More than 300 million people a week bring health questions to ChatGPT

OpenAI's Karan Singhal gave the figure on the a16z main feed, alongside how medical expertise is trained into a general purpose model and evaluated with hundreds of medical experts. The conversation covers Epic integrations, clinician copilots and personal health agents, and his argument that the cost of medical intelligence could fall sharply. The number matters independently of the thesis: whatever the quality of the answers, a consumer chatbot is now one of the largest health information channels in the world. The interviewer is an a16z partner investing in this category.

Source: The a16z Podcast, 20 October 2026, What Happens When Everyone Has a Health Expert in Their Pocket?

A bipartisan bill would put immersion DUV lithography under countrywide China controls

The MATCH Act sits in the Senate version of the NDAA and would shift immersion DUV lithography and a handful of other top tier tools from entity based to countrywide controls, directing Commerce to negotiate allied alignment for about 150 days before permitting unilateral action. Nick Brown of the Center for Technology and Statecraft called the gap a cannonball sized hole: China imported roughly 90 immersion DUV systems a year in 2024 and 2025, more than 10 billion dollars worth and close to two thirds of ASML's immersion DUV exports, and now holds about 340 ASML systems. The NXT:1980i currently permitted into China is sufficient for 7nm logic and HBM2E, and a single machine can produce over 100,000 Ascend chips a year even at Chinese yields. The real chokepoint he identifies is narrower still, the excimer laser light source, where there are only two suppliers worldwide, Cymer in the US and Gigaphoton in Japan.

Source: ChinaTalk, 20 October 2026, The Lithography Lever

Intel's CTO says tool supply, not capital, is what caps AI compute growth

AI compute demand has grown roughly three times a year for a decade, but Pushkar Ranade told Dwarkesh Patel that existing equipment capacity could roughly double leading edge output, not multiply it tenfold, and that ASML produces fewer than 100 EUV machines a year. Intel is betting on High NA EUV, which lifts numerical aperture from 0.33 to 0.55 for about 1.7 times smaller features per pass, arguing it pays off because Intel's backside power delivery relaxes front side pitch requirements in a way TSMC's process does not; machine cost estimates on the episode ranged from 200 to 400 million dollars. Intel is also helping bring up Elon Musk's Terafab, with Musk quoted saying that without it we do not have the chips. Ranade conceded Intel should have invested more and that selling its DRAM and flash businesses looks poor in hindsight, while claiming the company is within a few percent of technology parity and short on scale and execution.

Source: Dwarkesh Podcast, 22 October 2026, What will computer chips look like in 2045?

Modal has crossed 1 billion dollars annualised and raised 750 million at a 16 billion valuation

Erik Bernhardsson said the Series D follows a 5 billion dollar Series C, with nearly three billion sandboxes spun up to date, headcount around 200 heading to roughly 250 by year end. Inference functions are about two thirds of revenue and sandboxes, which the team once discussed shutting down, are now about a third, pulled along by vibe coding companies, background agents and reinforcement learning customers. His thesis is that GPU capacity is bought as a rectangle while demand is nothing like one, so the business is pooling demand across thousands of customers to smooth the load, now across about 25 clouds on mostly three year deals. The striking admission is which way the risk runs: those multi year reservations leave Modal effectively long GPU prices, so he worries more about prices collapsing than rising.

Source: Latent Space, 22 October 2026, Why AI Compute Comes in Rectangles, Erik Bernhardsson of Modal

IBM's chief executive expects open weight models to take most enterprise token spend

Arvind Krishna put the figure at 75 to 90 percent, on an economic gap he described as roughly an order of magnitude or more between frontier cloud API pricing and an optimised open weight model self hosted on private GPUs. Beyond cost he named protecting intellectual property and chain of thought from frontier vendors, plus geopolitics, and put the gap between Chinese and Western open weight models at six to twelve months. Palo Alto Networks chief executive Nikesh Arora gave the practical version: let staff use the best model while experimenting, then move production workloads to smaller task specific models, because you do not need a Ferrari for every task. Both guests sell into the shift they are describing, so weigh the claim accordingly, and note that specific figures from this episode come from an unofficial transcript without speaker labels.

Source: All-In, 21 October 2026, Arvind Krishna and Nikesh Arora: Quantum Threat, Open vs Frontier AI, AI Cyber Warfare and SaaS Collapse

Krishna puts cryptographically relevant quantum computing in the 2028 to 2033 window

His reasoning is that error correction on a running machine gets demonstrated around 2027, and gate operation counts scale roughly an order of magnitude a year from about 10,000 toward 10 million by 2029. The operative point is harvest now and decrypt later, meaning data stolen today becomes readable once the machines arrive, so post quantum migration has to start well before the break. He is dismissive of the idea that AI alone cracks encryption, since quantum algorithms need quantum hardware to beat classical methods. IBM sells both quantum hardware and post quantum migration services, which is worth holding in mind when reading the timeline.

Source: All-In, 21 October 2026, Arvind Krishna and Nikesh Arora: Quantum Threat, Open vs Frontier AI, AI Cyber Warfare and SaaS Collapse

The CFTC has opened public comment on futures contracts for compute

The episode reports the regulator taking comment on compute futures, which would turn GPU time into a tradeable, hedgeable instrument rather than something negotiated contract by contract. It is an early step rather than a decision, but it matters for anyone watching AI infrastructure get financialised: a liquid forward curve changes who can take capacity risk and how the cost of compute shows up on a balance sheet. The same episode reports Sergey Brin spending an unprecedented 100 million dollars this year opposing a proposed California billionaire tax. Note for the record that POLITICO Tech has been rebranded as The Decoded Podcast on the same feed, so the show is active rather than retired.

Source: The Decoded Podcast, 20 October 2026, Sergey Brin's huge spending spree to defeat the billionaire tax

Subscribe to our updates
Past daily issues Weekly newsletter