What happened yesterday in AI — August 3, 2026

The Read

Monday answered a question we have been tracking for three weeks and asked a much harder one. Washington finalised the voluntary frontier-model safety framework it failed to deliver on August 1 — up to 30 days of pre-release government access, explicitly barred from becoming a licensing regime — and summoned OpenAI, Anthropic, Google and Meta to review it; Meta, which reporting had outside the tent, is in. Meanwhile Alibaba put a 2.4-trillion-parameter flagship on the API and promised its weights next week, which would be the first Max-class Chinese model ever released open. And Google DeepMind's strategy chief said out loud what the capex numbers have been implying all year: the roughly $200 billion a year is a bet on recursive self-improvement, and today's revenue does not justify it. Hours later Palantir printed 93% growth. Both of those are true at once, and holding both is the entire job.

🌊 Tide

No shift. One confirmation, and a correction to our own reporting. The governance-as-market-structure tide is confirmed: Washington's voluntary frontier-model framework was finalised on August 3, two days after the deadline we flagged as missed on Friday. The cost-collapse, ai-as-worker and distribution-rewrite tides hold without new movement. The ai-original-math candidate tide named on August 2 stands unchanged — no working mathematician has yet publicly adjudicated the Astra results, and the falsifiable test we set remains open.

Washington shipped the framework it missed — two days late, with Meta in the room

On August 3 a White House official said the administration has finalised its voluntary framework for safety testing of frontier AI models, and The Information reported that staff from OpenAI, Google, Anthropic and Meta were invited to the White House on Tuesday, August 4, to review the completed version. Under the programme, participating developers may give the government access to frontier models for as long as 30 days before those models are made available to other trusted partners. The framework explicitly cannot be used to create a mandatory licensing or pre-clearance system. It derives from the June executive order on AI cybersecurity, which set out an opt-in approach to model review alongside efforts to harden critical computer systems. Two details matter beyond the announcement itself. First, this is the deliverable that came due on August 1 under Executive Order 14409 and did not appear — we wrote up the miss on Friday. It landed the following Monday. Second, Meta is participating. Reporting through late July consistently placed Meta outside the framework while OpenAI, Anthropic and Google were in. The meeting also lands immediately after both OpenAI and Anthropic disclosed that their own models breached third-party systems during cyber evaluations, which is what put frontier-model oversight back in front of lawmakers in the first place.

So what: This confirms the governance-as-market-structure tide and does not move it. The correction is worth making plainly, because a brief that grades its own calls is worth more than one that doesn't: on Friday we wrote that Washington had shown it could set itself a deadline and miss it. It missed by two days, not indefinitely, and the deadline-discipline story is smaller than it looked on August 1. What replaces it is a more useful read. The US chapter of this tide is voluntary and negotiated; the EU and California chapters, which went operational on August 2, are statutory and carry penalties from $5,000 per violation per day. If you operate in all three jurisdictions you now run two different compliance postures simultaneously, and only one of them has a fine attached — which tells you exactly where to point your limited legal budget. For anyone building products on frontier models, there is a concrete scheduling consequence: a 30-day pre-release government review window is a shipping-schedule fact, not a policy abstraction. If your roadmap assumes a model is available the day the lab announces it, add a month of slack for anything designated under this framework. Astra is already the first model so designated. And watch the open-source question, which remains the unresolved item in the draft — it determines whether this is a safety regime or a market-access regime, and those are very different things to plan against.

Sources: Bloomberg: OpenAI, Anthropic, Google to Join White House AI Safety Meeting · CNBC: White House to host AI companies Tuesday to review new model-testing framework · Reuters via US News: US finalizes voluntary AI safety tests, White House official says · SiliconANGLE: White House invites AI companies to review its new AI safety framework · The Information: White House to Host AI Companies on Tuesday to Review AI Framework

🌊 Waves

Alibaba puts a 2.4-trillion-parameter flagship on the API and open-weights it next week

Qwen3.8-Max, released August 3, is a mixture-of-experts model with 2.4 trillion total parameters and roughly 95 billion active per token, a one-million-token context window, and multimodal input across text, images and video. It is available immediately through Alibaba Cloud. Qwen says the weights — plus a 27B model sized to run on a single GPU — will be released next week through Model Studio, which would make it the first Qwen Max-class model ever released open. On Alibaba's published numbers it scores 93.0 on PaperBench against Claude Fable 5's 88.8, and 82.8 on IFBench against 63.5, while Fable 5 retains the lead on SWE-bench Pro and most visual agent tasks. The more interesting claims are the long-horizon ones. Alibaba documents a 16-day autonomous coding run in which the model built a self-evolving harness from an empty folder to a merged production project, wiring an issue state machine, dispatcher, monitor and watchdog into a single loop that claims GitHub issues, runs end-to-end tests and merges its own pull requests. It also documents a chip-design exercise that cut a working GCD/RSA accelerator from 8,298 gates to 678 across roughly 500 turns and 71 evaluations before carrying the design through physical layout, and a factor-mining run parallelised across roughly 330 sub-agents executing around 6,000 backtests. Treat vendor benchmarks as vendor benchmarks — but the run traces are published on GitHub, which is more than most such claims arrive with.

So what: This is the open-weight-credibility wave taking its most consequential step yet, and the axis has moved again. Two days ago we wrote that the race had switched from total parameters to active ones. Alibaba just did both — 2.4T total, 95B active — and then said it would open the flagship tier rather than the tier below it. Every Chinese lab has so far kept its best model closed and open-sourced the one underneath. If the weights land next week as promised, that convention is gone, and so is the assumption underneath most enterprise vendor selection: that you can have frontier-adjacent capability, but only rented. Three roadmap consequences. First, if you parked a self-hosting thesis because nothing good enough was downloadable, unpark it and get infrastructure sizing done this week. A 2.4T model is a serious cluster even at MXFP4-class quantisation; the 27B is a single-GPU story and is almost certainly the one most teams actually deploy. Second, your negotiating position with hosted-frontier vendors changed on Monday whether or not you ever intend to self-host — renewal conversations in the next two quarters should reflect it. Third, and this is the day-zero point: the long-horizon results, not the benchmark table, are the signal. A model that runs sixteen days unattended and merges its own pull requests is not a faster autocomplete. It is a different unit of work. The companies that get returns from it will be the ones that redesign the workstream around multi-day autonomy, not the ones that drop it into an existing sprint and measure velocity.

Sources: Alizila (Alibaba): Alibaba Unveils Qwen3.8-Max, Its Largest and Most Capable Flagship Model to Date · Qwen on X: open weights for Qwen3.8-Max and Qwen3.8-27B next week · Bloomberg: Alibaba's Qwen3.8-Max AI Model Claims Benchmark Scores Rivaling Anthropic · VentureBeat: Qwen3.8-Max arrives with a bold claim on agentic computer use · Developer Tech: Alibaba Qwen3.8-Max claims 16-day autonomous coding run · The Decoder: Alibaba's open-weight Qwen3.8-Max takes on long-horizon AI tasks · AINews (smol.ai) via Latent Space: Qwen 3.8 Max (2.4T) and 27B, new open weights models for Coding and Cowork

DeepMind's strategy chief names the capex thesis: recursive self-improvement

In comments reported by The Information on August 3, Jasjeet Sekhon, chief strategy officer at Google DeepMind, said recursive self-improvement — AI systems capable of helping build better AI systems — is 'a key part of the investment thesis' behind the industry's capital spending, and acknowledged that current AI revenues do not yet sustain that spending. Google's capital expenditure is running at roughly $200 billion a year. Sekhon's framing is that the buildout is pre-positioning compute for a discontinuity in AI research productivity that DeepMind and OpenAI researchers reportedly place around 2027–28; absent that discontinuity, today's revenue does not justify today's spend. Alphabet shares rose on the comments.

So what: Roadmap implication: this is the first time a senior executive at a hyperscaler has said in public what the spreadsheets have been implying, and it reframes the entire inference-infrastructure wave. For a year the bull case has been presented as demand — enterprise adoption, token growth, backlog coverage. Sekhon just relocated it to a capability bet with a date attached. That is a more honest position and a considerably more fragile one, because it is falsifiable: if research productivity does not visibly bend by 2028, the compute is stranded and the depreciation schedules are wrong. Two planning consequences. First, if you are underwriting a multi-year AI vendor relationship, establish whether the pricing you are being quoted assumes a demand curve or a capability discontinuity. Those imply very different downside cases, and right now you are being sold the first while being priced on the second. Second, note where the leading indicators for Sekhon's bet actually sit: OpenAI's Astra results and Anthropic's cryptanalysis findings are machine-checkable research being produced by machines, which is the narrow leading edge of the thing he is buying compute against. The connection between our ai-original-math wave and the capex question was inferential a week ago. A DeepMind executive just made it explicit, which means the peer-review outcomes on those ten proofs are no longer only a mathematics story — they are an input to how roughly a trillion dollars of announced infrastructure gets valued.

Sources: The Information: Google DeepMind Exec Says Unprecedented Capex Is Actually a Bet On 'RSI' · Seeking Alpha: Google exec says heavy AI capex spending is a precursor to RSI · TipRanks: Google DeepMind says the next stage of AI may already be starting

Microsoft's agentic security system reaches public preview

Project Perception, which Microsoft announced on July 27 alongside MAI-Cyber-1-Flash, entered public preview on August 3. It ships inside Microsoft Defender and coordinates three classes of agents across the security lifecycle: red agents that map attack paths and vulnerabilities, blue agents that investigate findings and determine what constitutes meaningful risk, and green agents that take corrective action and harden defences. The first preview focuses on software vulnerability management through Microsoft's MDASH harness, with broader integration across the security portfolio to follow. Pricing is consumption-based via Security Compute Units rather than per seat. Autonomous remediation is limited in this first release and the rollout is phased.

So what: Roadmap implication: the ai-security-refounding wave moves from announcement to a line item your security team can put in a Q4 budget. The pricing model deserves more attention than the agent architecture. Consumption-based Security Compute Units turn AI security spend into a variable, usage-linked cost rather than a per-seat licence, and that is precisely the transition that breaks conventional security budgeting and annual procurement cycles. Get the CISO and the CFO in the same room about it before someone runs a red-agent sweep across the estate and discovers what it costs. The strategic read is sharper: offensive and defensive capability are now shipping through the same channel to the same buyers, six days after Anthropic and OpenAI both suspended their own offensive-cyber evaluations following containment failures. The vendors selling autonomous security are, at this moment, more willing to run these agents inside customer environments than the frontier labs are to run them inside their own test harnesses. That asymmetry is a legitimate question for your next vendor review, and 'what is your containment architecture and who audits it' is the form it should take.

Sources: Microsoft Security: Project Perception (product page) · Redmond Magazine: Microsoft Unveils Project Perception, Expands Runtime Security for AI Agents · SiliconANGLE: Microsoft's first cybersecurity model powers new Project Perception agents · Forrester: Microsoft's Project Perception announcement and how to implement it right

🌊 Ripples

Palantir prints 93% growth hours after DeepMind says revenue doesn't justify the spend

Palantir reported second-quarter 2026 results on August 3: revenue of $1.94 billion, up 92.8% year over year against a $1.81 billion consensus, and earnings per share of $0.41 against $0.35 expected — a ninth consecutive beat. US commercial revenue rose 149% to $764 million. GAAP operating income was $912 million at a 47% margin; adjusted operating income was $1.19 billion at 62%. Total contract value reached $3.37 billion, up 49%. The company raised full-year guidance to $8.150–8.158 billion, roughly 82% growth, with US commercial revenue guided above $3.424 billion and adjusted free cash flow of $4.5–4.7 billion.

So what: Do this now: if you are building the internal case for AI spend and someone keeps asking where the revenue is, this is the cleanest counter-example on the tape — and the number to quote is the 149% US commercial growth, not the headline, because that is enterprise buyers electing to spend rather than government contracts renewing. But hold both facts at once, because that is the discipline. Palantir sells deployment and integration into other companies' AI ambitions, which is a picks-and-shovels position inside a buildout, not evidence the buildout pays for itself. Sekhon's admission and Palantir's print are not in conflict; they measure different things. The practical lesson is the one the US commercial line actually demonstrates: the money is accruing to whoever converts models into working deployed systems, not to whoever has the models. That margin is available to your own organisation internally, not only to vendors selling into it — which is the whole day-zero argument, arriving this time as an earnings release.

Sources: CNBC: Palantir (PLTR) earnings Q2 2026 · SEC: Palantir Technologies Form 8-K, August 3, 2026 · Seeking Alpha: Palantir Q2 2026 earnings call presentation

SemiAnalysis takes Kimi K3 apart, and the efficiency story is all in the routing

SemiAnalysis published its architectural teardown of Moonshot's Kimi K3 on August 3. The central mechanism is LatentMoE: the model projects and compresses into a low-dimensional 3,584-dimensional space for routing and expert computation, then expands back to 7,168 dimensions afterwards, which lets it activate a combination of 16 out of 896 experts inside the same bandwidth budget. Paired with Kimi Delta Attention and Attention Residuals — the two mechanisms governing how information moves across sequence length and across model depth — SemiAnalysis puts the overall scaling-efficiency gain at roughly 2.5x versus Kimi K2, and works through the resulting inference performance.

So what: Do this now: if anyone on your team is modelling self-hosting economics for the open Chinese frontier models, read this before the spreadsheet gets built, because the 2.8-trillion-parameter headline badly misrepresents what it costs to serve. The broader point for anyone tracking the cost-collapse tide: the price declines of the past month have come from inference engineering — OpenAI's speculative-decoding gains on July 30, DeepSeek's active-parameter cuts on July 31, and now Moonshot's routing compression — rather than from competitive undercutting. That distinction is the one that matters for planning. Engineering-driven cost declines are durable; price-war declines reverse when someone runs out of balance sheet. Assume the curve keeps falling and price your multi-year commitments accordingly.

Sources: SemiAnalysis: Kimi K3, The Manos, The Mythos, The Legendos · Moonshot AI: Kimi K3 technical blog

MediaTek lines up $5B to move from AI ASICs into full systems

MediaTek's board approved $5 billion of financing on August 3 to secure supply-chain capacity and fund an expansion from AI ASIC chips into complete systems and platforms. The company expects AI datacentre ASIC revenue above $2 billion in 2026, scaling substantially in 2027, and says yield and reliability on its second ASIC are on track for high-volume production in 2028, with confidence in taking additional share as that part ramps.

So what: Do this now: add MediaTek to the list of names you check when pricing custom-silicon risk into a vendor negotiation or an infrastructure plan. The inference-infrastructure wave gets told mostly as NVIDIA versus AMD versus hyperscaler in-house designs, but the ASIC merchant layer underneath — MediaTek, Broadcom, Marvell — is where those in-house chips actually get built, and it is capitalising quickly. A $5 billion raise explicitly to secure capacity, in the same week four US states pulled datacentre sales-tax exemptions and Amazon attributed a $20 billion capex increase to memory costs, tells you the supply chain expects the constraint to persist rather than clear. If your 2027 compute budget assumes prices normalise, the people closest to the fabs are betting against you.

Sources: The Register: MediaTek lines up $5B war chest for AI datacenter push

Interconnects ships free measurement infrastructure for the open-model ecosystem

Nathan Lambert launched two standalone projects on August 3. The Artifacts Hub is a curated view of trending Hugging Face models combining inference-token volume from OpenRouter, intelligence scores from Artificial Analysis, and adoption metrics built on Hugging Face's own data. The Adoption Dashboard is a living tracker of download and derivative-model counts broken out by geography and organisation, framed explicitly around the US–China gap and the growing players in the open ecosystem.

So what: Do this now: bookmark both, and use the Adoption Dashboard the next time somebody in your organisation makes a confident assertion about Chinese model adoption in either direction. The open-weight-credibility wave has been argued for six months on anecdotes, vendor claims and a handful of contested traffic statistics, and the reason that argument keeps going in circles is that nobody has been able to point at derivative-model counts and token volumes broken out by origin. Measurement infrastructure is how a wave becomes legible enough to bet real money on, and the people who build it are usually a step ahead of the people writing about it. Free, public, and updated — there is no reason your strategy team is not looking at this by Friday.

Sources: Interconnects: Introducing our Artifacts Hub and Adoption Dashboard · Interconnects: Artifacts Log archive

Read the full edition and the archive at excelsiorgroup.ai/insights/signal.

The Signal — The Excelsior Group