BGAD Consulting
STRATEGIES. DELIVERED.
BGAD News Flash
Daily briefing: digital, tech and AI
15 August 2026
Six Anthropic backers told the Financial Times they expect a float as early as October at more than $2 trillion, which would eclipse SpaceX's $1.77 trillion June debut and set an all time record. That is more than double the roughly $965 billion private valuation set at the company's last round, and investors project annualised revenue of $100 billion to $120 billion by year end against $47 billion in May. The detail worth holding onto is that the number comes from investors, not the company: senior Anthropic executives have reportedly not fixed a valuation target even privately. The same investors concede revenue growth slowed measurably in June, the month a US government order forced a global shutdown of the company's two most advanced models, and the underlying export control remains active.
Source: All-In, 14 Aug 2026, Anthropic's $2T IPO, Zuck's AI Manifesto, Nvidia's $500B AI Bet, Grok's Comeback. The hosts and guest Gavin Baker are active investors with AI exposure, so treat the bull case accordingly.
Reuters reported preliminary talks for Silver Lake to take Workday private at roughly its $43 billion market value, which would rank among the largest software buyouts on record. Workday shares jumped 16 per cent to $204.64 on the news, and Silver Lake, which manages around $114 billion, may bring in co investors. The reason this is more than a big cheque is the thesis behind it: analysts read the approach as the first concrete evidence that AI is compressing seat based SaaS multiples far enough for private equity to start picking off incumbents. Workday's appeal is a sticky ERP business, a deep HR data set and margin headroom, with one analyst arguing that going private would enable a more aggressive AI transition than public market scrutiny allows.
Source: All-In, 14 Aug 2026, Anthropic's $2T IPO, Zuck's AI Manifesto, Nvidia's $500B AI Bet, Grok's Comeback
Kevin Roose and Casey Newton picked apart Anthropic's disclosure that an unreleased research version of Claude improved the proven lower bound for the share of nontrivial Riemann zeta zeros lying on the critical line, from 41.6 to 67.2 per cent. The run burned 31 million output tokens, discarded around 650 failed ideas, and spent about 36 hours coordinating roughly 60 subagents inside Claude Code across 2,400 shell commands, with 2 agents finding the winning idea, 13 contributing pieces, 30 failing and 13 acting as validators. The insight was cross domain rather than brute force: combining 1973 techniques with a 2000 Bombieri paper in a way no mathematician had tried. It is emphatically not a proof of the hypothesis, which needs 100 per cent, but the result was machine verified in Lean and reviewed by outside number theorists, which makes it a rare checkable instance of a model producing novel research mathematics rather than a benchmark score.
Source: Hard Fork, 14 Aug 2026, Zuckerberg's Anti-Doom Fantasy, Finally an A.I. Detector That Works, A.I. Math
Bloomberg reported on 13 August that Anthropic is negotiating to acquire Decart AI for roughly $6 billion, which would be its largest known acquisition. Decart, founded in 2023 by brothers Dean and Orian Leitersdorf with Moshe Shalev, builds software that makes chips run more efficiently and cuts the cost of training and serving models. The price is about a 50 per cent premium to the nearly $4 billion valuation Decart reached in a May round, itself up from $3.1 billion in August 2025. The strategic read is that Anthropic is trying to extract more served demand from compute it already holds rather than simply buying more of it, which is a telling priority immediately before a listing.
Source: All-In, 14 Aug 2026, Anthropic's $2T IPO, Zuck's AI Manifesto, Nvidia's $500B AI Bet, Grok's Comeback
Google released Gemini 3.7 Flash, tuned for throughput rather than peak intelligence, at 340 tokens per second, more than twice the rate of GPT-5.6 Luna, with prices halved. Cognition said it matches Sonnet 5 coding performance for less than half the cost and Vercel's Brandon Galang called it Pareto optimal, though Artificial Analysis measured 40 cents per task, roughly eight times ultra cheap options, leaving it awkwardly placed on cost per unit of intelligence. The next day OpenAI previewed an Ultrafast mode running GPT-5.6 Sol on Cerebras hardware at about 750 tokens per second, API only for selected customers, aimed squarely at real time voice, support, commerce and security response. The pattern to note is that the model race is fragmenting into parallel races for distribution, harness quality and now latency, rather than one leaderboard.
Source: The AI Daily Brief, 14 Aug 2026, How to Decide What Work AI Should Do for You, The AI Deputization Audit
xAI shipped Grok 4.6 on 12 August, about five weeks after 4.5, live the same afternoon in Cursor, Grok Build and the API. It reuses the same 1.5 trillion parameter V9 foundation, with gains coming from improved supervised fine tuning and reinforcement learning aimed at long running agents and more ambitious visual work. It scored 61 on the Artificial Analysis Intelligence Index, up five points, overtaking Kimi K3 and matching GPT-5.6 Sol for third best in the world, while pricing held flat at $2 and $6 per million input and output tokens against Sol's $5 and $30. A larger 2.1 trillion parameter Grok 4.7 is expected within weeks.
Source: The AI Daily Brief, 13 Aug 2026, Grok 4.6 Shows How Fast Your AI Options Are Expanding; All-In, 14 Aug 2026. Note that several All-In hosts have disclosed investments in Musk companies.
AlphaSense ran US and Chinese models across several hundred financial analysis tasks and found the savings do not survive contact with real work: GPT-5.6 Sol came out 13 per cent cheaper and 20 per cent higher quality than Kimi K3, and Opus 4.8 beat both Chinese models at comparable cost. Chief executive Jack Kokko's explanation is that a higher price per token can still be cheaper overall when the model uses tokens more efficiently. The same week undercut the price argument from another direction, as DeepSeek V4 Pro landed well short of its leaked benchmarks, scoring 53 on the Artificial Analysis Intelligence Index against leaked claims of 87.9 per cent on Terminal Bench 2.1, with early testers calling it benchmaxxed. V4 Pro still costs roughly one twelfth of Fable, so the buying question becomes cost per completed task rather than sticker price per million tokens.
Source: The AI Daily Brief, 13 and 14 Aug 2026, Grok 4.6 Shows How Fast Your AI Options Are Expanding and The AI Deputization Audit
ChinaTalk analysed leaked minutes from a four hour meeting in late July between DeepSeek chief executive Liang Wenfeng and investors. Liang said the lab's largest revenue source is enterprises paying for API tokens, with enterprise revenue this year in the hundreds of millions of US dollars, enough that DeepSeek could turn a net profit and move toward a listing. He also said Huawei allocated DeepSeek 16,000 Ascend 950 GPUs, fewer than it sold to larger Chinese internet firms, arguing they need Huawei silicon more because DeepSeek can still "acquire some noncompliant chips". On research direction he called continuous learning the only problem worth solving on the path to AGI and dismissed world models as irrelevant to raising machine intelligence. Liang was reportedly furious once the minutes spread, pausing a funding round and pushing back IPO plans again.
Source: ChinaTalk, 13 Aug 2026, The DeepSeek Thesis by Irene Zhang. This is a written analysis piece rather than a podcast episode; ChinaTalk's latest audio is a music show.
Ed Elson and Semafor's Rohan Goswami dug into CME Group's announcement with Silicon Data of two compute futures contracts launching 5 October, subject to regulatory approval. The contracts are pegged to rental costs for Nvidia H100 and Blackwell B200 GPUs, with each contract representing a one month H100 rental, letting AI firms and investors hedge compute the way they hedge power or crude. Silicon Data chief executive Carmen Lee argues buyers of identical compute have been paying wildly different prices with no way to compare terms. The open question the segment leaves is whether standardising compute is genuine price discovery or another derivative layer stacked on an already leveraged buildout.
Source: Prof G Markets, 13 Aug 2026, AI Compute Futures: Has Wall Street Gone Too Far?
Brad Lightcap announced on 11 August that he was leaving to start something new, and chief revenue officer Denise Dresser followed on 13 August, less than a year into the role. They join a months long run of departures including Fidji Simo as chief executive of Applications, chief product officer Kevin Weil, Sora head Bill Peebles and business applications CTO Srinivas Narayanan. Elson's argument is that this is a market signal rather than gossip: losing commercial leadership immediately before an expected landmark listing is the kind of thing public investors should price in.
Source: Prof G Markets, 13 Aug 2026, AI Compute Futures: Has Wall Street Gone Too Far?
Bloomberg reports the coding agent company Cognition is in early talks for a round at a $40 billion valuation, three months after raising $1 billion at $26 billion, and is seeking roughly another billion. Sources say it has doubled its revenue run rate to $1 billion since that last raise. The comparison that gives the number meaning is Cursor, which sought $50 billion on $2 billion in annualised revenue in March before SpaceX bought it in a $60 billion all stock deal. If Cognition prices at $40 billion on half the revenue, that purchase starts to look like a bargain.
Source: The AI Daily Brief, 13 Aug 2026, Grok 4.6 Shows How Fast Your AI Options Are Expanding
CoreWeave's quarterly revenue doubled year on year to $2.6 billion while cash burn also doubled, to $5.7 billion, but investors fixed on a $104 billion backlog that had already grown $25 billion since the books closed at the end of June. Nebius posted 454 per cent revenue growth to $582 million, beat earnings per share forecasts by 83 per cent, and cleared its Blackwell compute auctions 15 per cent above its previous record price for Hopper. CoreWeave is up 19 per cent since reporting and Nebius up 34 per cent. Neoclouds absorb hyperscaler overflow, which makes them one of the cleaner read outs on marginal AI demand, and that demand held up in a quarter when token austerity was fashionable.
Source: The AI Daily Brief, 13 Aug 2026, Grok 4.6 Shows How Fast Your AI Options Are Expanding
Korean outlet ChoSun reports that the first three months of integrating Claude Code into Samsung's semiconductor software stack produced large gains, with complex jobs such as system on chip verification falling from three months to two days. In one case a second year engineer completed a task that normally takes a month in a single day. The report does not claim efficiency gains across the whole chip design process, and that qualifier is the interesting part. It is a good illustration of the jagged frontier of enterprise adoption, where the wins came from highly customised tasks and from letting a junior employee operate well above their experience level.
Source: The AI Daily Brief, 13 Aug 2026, Grok 4.6 Shows How Fast Your AI Options Are Expanding
A New Mexico court ordered Meta to pay $567 million into an abatement fund on top of the $375 million levied in March, taking the total to roughly $942 million. The 68 page ruling found Meta built features designed to optimise engagement that are harmful to teenagers and failed to adequately warn users, with the money split across treatment, screening, prevention, referrals and oversight. More consequential than the cash are the product orders: Like counts hidden from under 18s without parental approval, push notifications to minors paused between 10pm and 7am, and usage capped at 90 hours a month. Court mandated product design is a materially different remedy from a fine, and worth watching as a template.
Source: Hard Fork, 14 Aug 2026, Zuckerberg's Anti-Doom Fantasy, Finally an A.I. Detector That Works, A.I. Math
Alongside Zuckerberg's essay, Meta confirmed it will release open weights for Muse Spark 1.2, its latest and most capable model, plus a new open source family called Muse Glimmer designed to run locally on a laptop. Coverage framed it as a direct swipe at OpenAI and Anthropic, both of which keep frontier weights closed. Roose and Newton questioned whether the openness pitch is credible given Meta had drifted away from open releases at the frontier over the past two years. For enterprises weighing self hosting, a genuinely open frontier grade model changes the build versus buy maths more than another API price cut does.
Source: Hard Fork, 14 Aug 2026, Zuckerberg's Anti-Doom Fantasy, Finally an A.I. Detector That Works, A.I. Math