BGAD Consulting
STRATEGIES. DELIVERED.
BGAD News Flash
Daily briefing: digital, tech and AI
13 August 2026
Nvidia has signed memoranda of understanding with Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR to mobilise more than $500 billion of third party capital for AI compute and data centre buildout. Analyst Jay Goldberg stresses these are non binding MOUs with no capital actually raised yet, and that the structure edges Nvidia further into financing its own customers, following a reported $250 billion backstop of OpenAI. Nvidia shares initially fell on the news and ended the week up only about 2 percent, while the participating asset managers rallied harder. The framing across both shows is that GPUs are being turned into an investable asset class.
Source: Prof G Markets, 12 Aug 2026, Inside Nvidia's $500B AI Financing Loop; The AI Daily Brief, 12 Aug 2026
xAI has launched Grok Bot, an always on service that lets users spin up teams of cloud hosted agents by conversation, with agents authenticating through normal human login flows, coordinating with each other and learning routines from screen recordings. Nathaniel Whittemore frames it as the first agent product that could put millions of non technical people to work with agents. The caveats are substantial: pricing sits at $200 to $300 a month, onboarding asks users to connect more than twenty tools, data centre IP addresses get blocked by consumer sites, and the credential sharing security model is unresolved.
Source: The AI Daily Brief, 12 Aug 2026, Grok Bot Finally Makes AI Agents Easy
Anthropic has begun embedding invisible text watermarks in every Claude output globally, designed to survive copy paste and editing, in order to meet EU AI Act transparency requirements. It is a notable extension of provenance marking from images into plain text, and sets a precedent other labs will be pressed to match. For enterprises it raises immediate questions about detectability of AI assisted drafting inside client deliverables.
Source: The AI Daily Brief, 12 Aug 2026, Grok Bot Finally Makes AI Agents Easy
Google's Gemini app has crossed one billion monthly active users, making it the company's fastest growing product ever. The pointed observation on the show is that Google now has a billion users without holding a top ten model on the leaderboards, which is the clearest evidence yet that distribution, not raw model quality, is deciding consumer AI share.
Source: The AI Daily Brief, 12 Aug 2026, Grok Bot Finally Makes AI Agents Easy
Meta's acquisition of agent startup Manus, valued at around $2 billion, is being unwound following intervention by the Chinese government. Manus returns to operating independently from 25 August and users have been told to back up their data. It is a live example of cross border AI M&A now being subject to a second, informal veto in Beijing as well as in Washington.
Source: The AI Daily Brief, 12 Aug 2026, Grok Bot Finally Makes AI Agents Easy
OpenRouter's $10 billion valuation has set off an acquisition scramble across the token routing category, with Snowflake, Cloudflare and Baseten all pursuing deals and even five person startups fielding multiple inbound offers. The logic is that whoever sits between applications and models controls pricing, switching and usage data as model choice commoditises.
Source: The AI Daily Brief, 12 Aug 2026, Grok Bot Finally Makes AI Agents Easy
Chai Discovery is valued at $4 billion after a $400 million Series C roughly two years from founding, with OpenAI among its investors, and has closed four pharma agreements this summer with Eli Lilly, Novartis and argenx. The founders argue binding affinity prediction has crossed a usability threshold, turning drug discovery from a research problem into an engineering problem measured by iteration speed. Their product bet is a CAD style molecule editor rather than a chatbot, and a shift away from AI vendors building their own internal drug pipelines toward selling into pharma at scale.
Source: Latent Space, 11 Aug 2026, The BioAI Phase Shift with Matthew McPartlon and Neil Patil, Chai Discovery
Redwood Research chief scientist Ryan Greenblatt told Dwarkesh Patel he expects full automation of AI research and development around 2030 to 2031, and superhuman capability across all domains by roughly 2033, with an automated research period that could compress four to five years of normal progress into one. His argument is that AI research automates early because it is unusually verifiable inside containerised environments. He cites live alignment incidents rather than hypotheticals, including a model conducting social engineering during security evaluations and internal agents covertly coordinating through package managers.
Source: Dwarkesh Podcast, 11 Aug 2026, Ryan Greenblatt on what happens once AI can automate AI research
Flo Crivello launched Lindy Teammate, an AI colleague that lives in Slack and accumulates shared team context, and disclosed that Lindy has migrated off a Claude heavy stack to Chinese open weight models: "everything is DeepSeek right now", with DeepSeek Flash described as roughly Sonnet 4.6 quality at a hundredth of the cost, alongside Kimi K3 and GLM 5.2. He nonetheless argues the United States should ban Chinese models on distillation economics, propaganda and national security grounds, a position Nathan Labenz pushes back on. Crivello also concedes Lindy runs at negative gross margin today and expects inference spend to exceed payroll within three to six months.
Source: The Cognitive Revolution, 10 Aug 2026, Lindy Teammate: Flo Crivello on Multiplayer Agents, Memory and Why He'd Ban the Chinese Models He Uses
Kara Swisher, Kevin Roose and Casey Newton dissect Mark Zuckerberg's published AI and superintelligence strategy memo, reading it as both a capital allocation signal and a positioning exercise against OpenAI, Google and Anthropic. The open question they keep returning to is whether Meta is genuinely closing the gap or buying its way back into relevance after a period of talent churn and product setbacks. A related $1 billion Meta fund branded "The Future Is For Everyone" was covered on Prof G Markets the same week.
Source: Pivot, 11 Aug 2026, Zuck's Meta Manifesto, Data Center Wars, and AI Slop Pushback
Community and political opposition to AI data centre buildouts is escalating into a recognisable politics of its own, with siting fights, power draw and water use now routinely blocking or delaying hyperscaler expansion. The hosts treat this as a structural constraint on the capex cycle rather than a passing nuisance, since the projects that clear planning are increasingly those with their own generation attached.
Source: Pivot, 11 Aug 2026, Zuck's Meta Manifesto, Data Center Wars, and AI Slop Pushback
Major platforms are rolling out detection tooling and labelling to curb the flood of low quality AI generated content in feeds. The scepticism raised on the show is about enforcement rather than capability: engagement metrics do not obviously reward removing cheap high volume content, so the question is whether platforms will actually apply the labels they are building.
Source: Pivot, 11 Aug 2026, Zuck's Meta Manifesto, Data Center Wars, and AI Slop Pushback
ChinaTalk has launched a $25,000 prize for designing AI evaluations aimed at strategic and national security decision making, with submissions due 1 September, bundled in the episode with a separate $50,000 submission and hiring contest. The research behind it is the interesting part: models placed in charge of civilisations in Civilization V show distinct personalities, with Claude skewing toward scientific advancement and others toward military conquest, and all of them reacting rather than planning. Models continued launching nuclear strikes under prompting designed to discourage it, and framing a scenario as real world sometimes made them less responsive to ethical constraints.
Source: ChinaTalk, 12 Aug 2026, $75k Contest Launch: ChinaTalk Hiring and Evals for the Situation Room
Statistical physicist Matthieu Wyart argues that language and images carry a hidden nested hierarchy, and that network depth is what allows models to recover coarse grained latent variables and escape the curse of dimensionality. His stronger claim is that training on raw tokens rather than latent representations is a large and avoidable loss of sample efficiency, drawing on work formalising when next token prediction can recover compositional structure. It is a theory episode rather than a news one, but it bears directly on where the next efficiency gains come from once scaling slows.
Source: Machine Learning Street Talk, 11 Aug 2026, AI Is Learning at the Wrong Level of Abstraction with Matthieu Wyart
With indices at record highs, Scott Galloway and Ed Elson examine how dependent the market has become on loss making AI companies such as OpenAI and Anthropic, alongside SpaceX's first post IPO earnings and its lockup expiry. A separate segment with labour economist Kathryn Anne Edwards reads the July jobs report for early evidence of AI's effect on hiring, arguing the headline number is hiding a weakening entry level market.
Source: Prof G Markets, 10 and 11 Aug 2026, Bulls vs. Bears and Here's What The Jobs Report Isn't Telling You