BGAD Consulting
STRATEGIES. DELIVERED.
BGAD News Flash
Daily briefing: digital, tech and AI
9 October 2026
John Foley, who runs Lex at the Financial Times, told Ed Elson that SpaceX is raising about 40 billion dollars, roughly 30 billion in investment grade bonds plus 10 billion in bank loans. That comes four months after an 86 billion dollar listing and a 25 billion dollar bond sale within two weeks of it. The money is for Nvidia chips and a plan to stand up 10 gigawatts of data centre capacity by the end of next year, which at roughly 50 billion dollars a gigawatt implies around 500 billion dollars of spend. SpaceX credit default swap spreads have widened to 194 basis points from 110 in June, and Foley argued the 2.3 trillion dollar valuation rests on faith in Musk and a small float, with Starlink the genuinely profitable part and data centre revenue still early.
Source: Prof G Markets, 8 October 2026, SpaceX Raised $111B, It Already Needs More
On the night of 6 October OpenAI published hundreds of proofs covering hard problems in number theory and algebraic geometry, among them a claimed advance on the Riemann hypothesis. None of it has been formally reviewed. Rutgers mathematician Alex Kontorovich said a human who did this would win a Fields Medal on the spot. Ed Elson called it a real push at the boundary of mathematics while criticising OpenAI for building on existing research without proper attribution, and noted the company is reportedly chasing a 1.5 trillion dollar valuation, about twice Walmart's on less than a tenth of the revenue.
Source: Prof G Markets, 8 October 2026, SpaceX Raised $111B, It Already Needs More
Reflection unveiled Beam, a text only mixture of experts model built for coding, agentic and scientific work, at roughly 501 billion total parameters with about 23 billion active, trained from scratch and released under Apache 2.0, with full weights due later this month. Reflection claims 80.9 on SWE-bench Verified and three to four times the inference efficiency of GLM 5.2. Independent commentators place it around GLM-5.2 level, which still leaves it behind the leading Chinese open models. Axios puts Reflection's compute bill at about 150 million dollars a month on Colossus plus a 1 billion dollar Nebius deal.
Source: The AI Daily Brief, 6 October 2026, Point-Counterpoint: Consumers Will Never Pay for AI
Positron founder Tom Sohmers told Nathan Labenz that token spend became the company's largest non manufacturing expense, briefly exceeding human salaries and peaking above 100,000 dollars a day, before Opus 5.5 brought the cost of many tasks down. The agents run closed loop chip testing on Cadence Palladium emulators at around 500 kilohertz, against roughly 10 hertz for RTL simulation. Positron's Asimov effort splits about 60 percent design and 40 percent verification, where Jensen Huang has put Nvidia at roughly 20 and 80. The company raised 875 million dollars at a 5 billion dollar valuation on 10 September.
Source: The Cognitive Revolution, 8 October 2026, AI:AM: A Level We Shouldn't Pass? Notes from The Curve + Tokens vs. Salaries & Is SaaS Cooked?
Labenz reported that a contributor going by Hyperstition nearly halved the NanoGPT speedrun record in a single step, a bigger gain than the previous 45 improvements combined. The author credits the core insight mostly to humans rather than to a model. Labenz ties that back to a claim he heard at The Curve in Berkeley, that current models may already have the research taste needed for paradigm level breakthroughs, with elicitation rather than raw capability as the bottleneck.
Source: The Cognitive Revolution, 8 October 2026, AI:AM: A Level We Shouldn't Pass? Notes from The Curve + Tokens vs. Salaries & Is SaaS Cooked?
Kevin Mandia, the Mandiant founder now running Armadin, told a16z's David George that the company has found more than 90 zero days in customer production environments since January, and that the most recent were found by AI agents rather than by people. In internal testing across 24 real world kill chains, open weight and closed models all found the same eight vulnerabilities, differing only in speed and cost, which Mandia reads as evidence that slowing model releases would now come far too late to matter. He expects continuous AI red teaming to replace penetration testing as a category. Note the investor position: a16z co led Armadin's 255.5 million dollar Series B at more than 2.5 billion dollars, announced five days before this episode went out.
Source: the a16z Podcast, 6 October 2026, Building Defense for the Agentic Era: Kevin Mandia
AWS chief executive Matt Garman told a16z that the company reserves scarce GPU capacity specifically for startups, and set out where its own Trainium and Graviton silicon sits in the AI stack. The episode puts Amazon's capital investment at 220 billion dollars and looks at how the bottlenecks in the infrastructure buildout are shifting as agents start writing code and managing infrastructure themselves. Garman also covers what enterprises still want in place before they will trust autonomous agents. This is a venture firm's own show, so read the framing as a16z positioning on the buildout as much as reporting on it.
Source: the a16z Podcast, 8 October 2026, Building the Cloud for an Agentic World, AWS CEO Matt Garman
Reuters reported in August that the FCC was drafting a ban on import authorisation for new Chinese optical transceiver models. Writing for ChinaTalk, Quinn Ennis argues the espionage case is thin, because hyperscalers encrypt traffic before it reaches the module and transceivers cannot read or originate data. Chinese firms dominate final assembly but depend on Western inputs: US and European DSPs, STMicroelectronics microcontrollers, most high speed lasers, and TSMC fabs. A ban would therefore land hardest on Marvell, Lumentum and Semtech, while China refines about 69 percent of global indium and added indium phosphide to its export control list in February 2025.
Source: ChinaTalk, 7 October 2026, China, the FCC, and the Logic of Transceivers
Liam Fedus and Ekin Dogus Cubuk argued on Latent Space that no amount of model scaling substitutes for testing conjectures in a physical lab, so Periodic builds its reinforcement learning environments from real lab data rather than from answers already published in papers. They put the choke point at characterisation, mainly X ray diffraction, which is where they are pointing AI first, and say density functional theory cannot capture microstructure, superconducting transition temperature or strongly correlated electrons. Fedus disclosed forward deployed engineers already embedded with semiconductor partners, running inference locally on partner data, and a shift from selling copilots towards pricing on scientific outcomes. Co-host Brandon said on air that the company is about to announce a large fundraise, without naming a figure.
Source: Latent Space, 8 October 2026, Synthesis Superintelligence: from Semiconductors to Superconductors
Nathaniel Whittemore built an episode around a single number: close to 98 percent of US households do not pay for an AI subscription. He runs both readings of it. Either this is the largest untapped consumer market in software, or it is evidence that Silicon Valley has spent three years building AI mainly for itself. His candidates for an eventual consumer business are power user spending, entertainment and advertising.
Source: The AI Daily Brief, 6 October 2026, Point-Counterpoint: Consumers Will Never Pay for AI
John Mowrey, chief investment officer at NFJ, told Ed Elson the index is up more than 14 percent this year and trading near 19 times forward earnings, below its five year average, with forward earnings growth near 38 percent and roughly 900 billion dollars of hyperscaler capital spending behind it. His line is that the market may be mistaking a bubble for a boom, and that earnings strength is broader than the handful of big tech names. Elson named circular financing as the concern he takes seriously. Mowrey agreed the open question is whether the buildout is funded from debt or from operating cash flow, and said the risk is not fully priced.
Source: Prof G Markets, 7 October 2026, Why The S&P Just Hit A Record High, Despite Soaring Yields
Constellation Energy shares rose more than 12 percent after Google agreed a 20 year nuclear power purchase agreement, reported in the show's market recap without further comment from the hosts. It is the same pattern showing up across the hyperscalers: data centre demand is increasingly being met with very long dated contracts for firm generation rather than bought on the spot market.
Source: Prof G Markets, 7 October 2026, Why The S&P Just Hit A Record High, Despite Soaring Yields
POLITICO's relaunched daily show opens with the line that the AI accelerationists are turning on OpenAI, after Sam Altman moved towards a more pro regulation position and split the coalition that had wanted faster development and lighter rules. The same episode covers testimony from AI industry leaders at a New York City Council hearing the day before. The Council sat as a committee of all 51 members on 5 October, and OpenAI, Google, Anthropic and Meta all appeared, two of them after initially declining and one shortly before subpoenas were due to issue.
Source: The Decoded Podcast, 6 October 2026, AI industry reacts to Sam Altman's comments, AI leaders testify in New York City
Shawn Wang told Nathan Labenz he has put two or three employees on performance review for shipping low quality AI output, and that he will not pay for what he calls LLM psychosis. His 10,000 dollar bounty to replace a subscription costing more than 40,000 dollars a year drew mostly weak submissions, so his company built its own tool instead. Labenz agreed that SaaS is quite cooked for CRUD applications. Wang's counter is that the software market is not a fixed pie, that custom software will expand into the long tail, and that compute, specifically CPU supply, is the constraint that comes up most in his interviews.
Source: The Cognitive Revolution, 8 October 2026, AI:AM: A Level We Shouldn't Pass? Notes from The Curve + Tokens vs. Salaries & Is SaaS Cooked?
Bharat Patel, who leads AI and data for Accenture's defense portfolio, told Jordan Schneider that data rather than models is the binding constraint, and that the department has no continuous machine learning data collection strategy. Operational data is routinely deleted or parked in storage where it loses the context that would make it trainable. He points to Project Maven, where early models underperformed because the imagery collected simply did not contain the relevant targets, and to an Army Research Lab team that had to build bespoke hardware to gather second generation FLIR imagery of tanks. His counter example is Ukraine, where autonomy grew out of years of quietly collecting and labelling battlefield data. The episode is sponsored by Accenture Federal Services and Patel's closing argument is that government will need integrators, so weigh it accordingly.
Source: ChinaTalk, 6 October 2026, Data is the Hard Part