❯ Hassabis Steps Down as Google DeepMind CEO; Jeff Dean Departs to Found Discovery Loop

[DOUBLE EXIT] Google announced two personnel moves on the same day: Hassabis is stepping down as Google DeepMind CEO, becoming chairman of the lab and Alphabet chief scientist, with former CTO Koray Kavukcuoglu taking over as senior vice president; chief scientist Jeff Dean is leaving after 27 years of service to start a company. Hassabis said on X that the new role lets him focus on long-term strategy and channel more energy back into drug-discovery company Isomorphic Labs.

[HANDOFF & EXIT] Kavukcuoglu had already been doubling as Google’s chief AI architect; with this move he goes directly from technical lead to the lab’s top job, keeping both titles. Jeff Dean joined Google in 1999 as employee No. 30, wrote MapReduce and Bigtable, co-founded Google Brain in 2011, and was the driving force behind the TPU project — all of Google’s large models today run on that self-developed chip line. His new company is called Discovery Loop, registered as a public-benefit corporation with one mission: automating machine learning, scientific research, and engineering research. Leaving with him are Sanjay Ghemawat, Oriol Vinyals, and Quoc Le — a quartet that is arguably the longest-running collaboration in Google’s research system. Google is serving as founding investor and cloud provider, with Radical Ventures and Khosla Ventures co-leading the seed round; the round has yet to close and the valuation is undisclosed.

[TALENT GATE] For recruiting leads at other labs, this is a door that has suddenly opened. Nathan Lambert of the Allen Institute for AI publicly put out the call on X that same day, saying anyone on the Gemini team who wants to work on open-source models can come to him for an introduction. What’s really being rewritten is the retention assumption for Google’s internal research talent: for the past decade, “working with Hassabis and Dean” was itself a reason to stay; now both reasons have been pulled away at once, while the successor’s role leans more toward product and engineering integration. Alphabet used a seed investment to keep Dean within its orbit — securing a collaboration channel while conceding that it could not retain him.

❯ Anthropic confirms first in-house chip team, designing custom silicon for Claude

[DISCLOSURE] Anthropic confirmed to Business Insider that it is building an in-house silicon team to design custom chips for Claude — its first public acknowledgment of that work. The company frames it as hardware-software “co-design,” with chip architecture tailored directly to Claude’s compute profile, while making clear it still follows a multi-chip path — AWS, Google, Nvidia, and AMD remain the backbone of its infrastructure. Per the report, one publicly posted engineering role offers an annual salary of $320,000 to $485,000 and explicitly requires a track record in volume semiconductor production.

[WHY NOW] This step is not about distancing itself from anyone. Anthropic already holds a full suite of external compute agreements, but the mismatch between external suppliers’ production scheduling and its own product iteration cadence is a bottleneck shared by every frontier lab — you can order cards, but you cannot order a card tuned to your model’s shape. The payoff of in-house silicon lands on inference cost, not training compute: Claude’s enterprise call volume doubles year over year, and the unit cost of serving will eat straight into gross margin. The reference point is right next door — Google spent a decade using TPUs to press inference cost into a range it controls, a line single-handedly pioneered by the just-departed Jeff Dean. What Anthropic is catching up on now is lesson one of that course.

[BUYERS ASK] Enterprise customers’ technology-selection checklists will gain one more column: whether the supplier’s inference-cost structure, two or three years out, is held in its own hands or someone else’s. That maps directly onto the discount headroom and service-level commitments in long-term contracts. For Nvidia and AMD, near-term orders are unaffected, but major customers have gone from pure buyers to half-peers, narrowing the information asymmetry at the negotiating table. First to feel the pressure are cloud vendors’ custom-chip businesses — their original pitch was “you don’t need to build your own chip.”

▪ SIGNAL As model companies push one layer deeper into silicon, the middlemen selling compute lose one layer of justification for markup.

❯ Meta debuts first coding agent Muse Code, with output priced at $4.25 per million tokens

[PRICING HOOK] Meta has introduced its first coding agent, Muse Code, now in beta — installable in a terminal with one command, able to plan changes, write code, and verify results on its own. Powering it is the same-day Muse Spark 1.2 release, with the two jointly trained. Pricing is usage-based: $1.25 per million input tokens and $4.25 per million output tokens, with cached input as low as $0.15. Meta’s chief AI officer, Alexandr Wang, told CNBC that the low entry price is the hook this time.

[RELEASE CADENCE] Muse Spark 1.2 posts 54 points on Artificial Analysis’s intelligence index — 3 above July’s 1.1 release and 11 above the April original, the third launch in four months, one of the densest release cadences among US labs and tied with SpaceXAI for third by score. Meta reports Muse Code at 59% on DeepSWE 1.1, ahead of Grok Build 4.5 and Gemini 3.6 Flash — a number that is currently vendor-reported only. Researcher teortaxes flags a more telling detail: Muse Spark’s input, output, and cache-hit prices are all below DeepSeek V4-Flash — undercut even the column where Chinese models have long been strongest. Meta also introduced a “contributor tier” priced more than ten times below pay-as-you-go, conditional on developers agreeing to have their session data used to improve the model.

[DATA FOR SHARE] Trading model gross margin for call data is the only explanation that makes sense of this pricing. Enterprise buyers see more than the unit price: the contributor tier, ten times cheaper, routes codebases, error reports, and fix trajectories to Meta — precisely the training material coding models lack the most. First in the line of fire are Anthropic’s and OpenAI’s coding product lines, whose pricing assumptions rest on coding as a high-value scenario that can be sold at a premium. For engineering leads, the trade-off to put on the table and settle is whether the call fees saved outweigh the compliance cost of handing internal code over.

▪ SIGNAL The coding-model price war is on — the first shot lands on competitors’ margins, and the ammunition is data.

❯ Microsoft Filings Show $24.1B in Revenue From OpenAI in Fiscal Year Ended June

[CONCENTRATION] Bloomberg, citing regulatory filings, reports that Microsoft recorded $24.1 billion in revenue from OpenAI in the fiscal year ended June. On that basis, this one customer contributed more than half — possibly close to 70% — of Microsoft’s total AI revenue. A Microsoft spokesperson confirmed the figure includes all sales and revenue sharing from OpenAI. Backed out from Microsoft’s previously disclosed AI business growth rate, the full-year AI revenue pool amounts to roughly $34 billion.

[SCOPE] This disclosure is the first to show what Microsoft’s “AI business” category actually contains. Microsoft’s definition of AI revenue is quite broad — it includes revenue from all AI customers as well as AI-specific products sold to any customer. Previously, the market assumed the number was supported by a wide rollout of Copilot subscriptions and Azure AI services, and investors applied a “multi-customer, sustainable” valuation logic. Now the filing lays out the structure: after removing OpenAI, the remainder is less than $10 billion. What’s more, money flows in both directions between Microsoft and OpenAI — Microsoft is OpenAI’s largest shareholder and cloud provider, and OpenAI pours much of the capital it raises back into Azure, creating a closed loop. That loop looks like synergy in a growth phase, but in a slowdown it is concentration risk.

[VALUATION] Analysts who use Microsoft’s AI revenue as a demand thermometer need to rework their models: they previously read the figure as a gauge of enterprise AI adoption; now it is closer to a proxy for OpenAI’s cloud spending. The most directly affected is Azure’s growth narrative — if OpenAI’s capex pace changes, Microsoft’s AI revenue curve will move with it. For Microsoft’s own sales organization, the pressure falls on the remaining sub-$10 billion: only if that business accelerates can AI revenue be detached from a single customer.

▪ SIGNAL Microsoft’s largest AI customer is also one it invested in itself — in a growth phase the ledger reads as synergy; otherwise, it is exposure.

❯ Musk says SpaceX will use only Nvidia chips, betting on Vera Rubin architecture

[EXCLUSIVE] On SpaceX’s earnings call, Musk said the company has decided to build “entirely on Nvidia”, because “we believe Vera Rubin is the best architecture.” He also said SpaceX would get a significant portion of Nvidia’s GPU supply next year, and plans to send space-optimized Vera Rubin NVL72 racks into orbit. Nvidia shares rose after the announcement.

[DATA CENTERS TO ORBIT] The scale of compute SpaceX wants is the premise behind this news: the company’s stated goal is to build 10-gigawatt-class AI compute by 2027; working backward from the power density of a single Rubin card, the number of GPUs needed would be more than two million. That scale would immediately hit two walls on Earth — grid access and cooling water — and those are exactly the core points of contention in states’ recent tightening of data center approvals. Musk’s solution is to run the same architecture on the ground and in orbit; the orbital version would use nodes made of several dozen Vera CPUs paired with Rubin GPUs. He himself also admitted that heat dissipation and long-term reliability in vacuum remain unsolved engineering problems. Worth comparing: his earlier stance at Tesla was lukewarm, saying Rubin would be hard to scale in the short term, while Tesla is building its own AI hardware — the same person has two chip strategies across two companies.

[SUPPLY RESHUFFLE] Nvidia’s allocation decision has directly rewritten the queue positions of other major customers: as SpaceX takes a significant slice of next year’s supply, cloud providers and labs will have to recalculate the amounts they receive. What AMD loses this round is not just orders, but a reference customer — Musk’s public statement that it’s “the best architecture” is more persuasive to buyers than any benchmark. And the business-model assumptions of the satellite internet and remote sensing industries are also being shaken: if orbital compute actually works, data won’t all have to be downlinked for processing, and bandwidth — a long-standing constraint — would carry less weight.

▪ SIGNAL A single sentence — “only Nvidia” — has already set aside a large chunk of next year’s GPU production schedule.

❯ DeepSeek Restarts Second Funding Round, Plans to Raise RMB 50 Billion at Pre-Money Valuation of ~RMB 500 Billion

[RESTART] Multiple dealmakers said DeepSeek has restarted its second funding round, planning to raise RMB 50 billion at a pre-money valuation of roughly RMB 500 billion, with signing targeted for late August. The round was initially launched in mid-July and abruptly halted at the end of July. Reports at the time tied the halt to founder Liang Wenfeng’s displeasure over a so-called “investor meeting minutes” circulating online. These details are based on dealmakers’ accounts; the company has not yet issued a public confirmation.

[VALUATION HIKE] The benchmark is the first round. According to earlier public reporting, DeepSeek began its first round in April 2024 and closed in June, also raising RMB 50 billion at a valuation above RMB 350 billion — the largest first-round financing in the history of Chinese large models. The same raise now commands a valuation roughly 40% higher. The supply-demand picture is even clearer in the first round’s subscription: capital expressing investment intent topped RMB 1 trillion, yet only RMB 50 billion was admitted. That leaves at least RMB 500 billion that didn’t get a ticket — and this restart gives it a second shot. Overseas observer poezhao adds a detail: the largest single check in the first round was from Liang Wenfeng himself, at about RMB 3 billion.

[ANCHOR RESET] For AI investors in China’s primary market, RMB 500 billion becomes the new pricing anchor: every large-model valuation negotiation from now on must first explain how it differs from DeepSeek. What this anchor presses down on is the funding rhythm of the second tier — the same pool of money now waits behind DeepSeek, leaving far less room on terms. A caveat: the signing date is the dealmakers’ version. July already showed how a round can be called off at any moment. Until the company officially announces, this is still a deal that hasn’t landed.

▪ SIGNAL RMB 1 trillion of intent chases an RMB 50 billion quota — money was never the scarce asset.

❯ CXMT Refuses Apple’s Price Cut, Demands Memory Quotes No Lower Than Samsung and SK Hynix

[LEVERAGE SHIFT] Apple reportedly entered talks with CXMT over supply pricing for LPDDR5X and other mobile memory in a bid to cut component costs on the next-generation iPhone — and its price-cut demands were rejected. CXMT insists on quoting no lower than Samsung Electronics and SK Hynix, in other words, it is done playing the role of the supplier that buys its way in with low prices.

[DOMESTIC ORDERS] Underpinning that stance is not capacity — it’s the order book. Domestic players such as Huawei and Xiaomi, seeking supply security, have locked away a large share of CXMT’s DRAM capacity in long-term contracts. With no shortage of buyers, the company has no reason to concede on price to Apple. The earnings picture backs the same reading: a Counterpoint report shows CXMT and Nanya Technology grew revenue 716% and 690% YoY last quarter, respectively, and analysts broadly view CXMT’s ascent to No. 4 in global DRAM as a foregone conclusion.

[COST MODEL] Apple’s supply-chain team has lost a price lever: it once wielded the Chinese memory maker as an alternative in negotiations with Samsung and Hynix — now that alternative quotes the same price tier itself. What gets pushed up is the component-cost floor for device makers — a floor previously propped up by the idea that “there’s always someone willing to be cheaper.” Domestic substitution has come this far, and price is the real watershed.

▪ SIGNAL The coming of age for domestic substitution: daring to say no to a big customer for the first time.

❯ Zhang Yiming Tells ByteDance AI All-Hands: No Distillation Shortcut to Boost Model Capabilities

[FOUNDER'S CALL] The Information reported, citing sources, that ByteDance founder Zhang Yiming told employees at the AI team’s all-hands meeting in July that the company will not use model distillation to accelerate capability gains — even if it means falling behind domestic peers in the short term. The gist of his remarks: willing to sacrifice some short-term gains for long-term goals.

[WHY BYTEDANCE] Distillation refers to training one’s own model on the outputs of a stronger frontier model; multiple Chinese companies in the industry have faced accusations over the practice, while ByteDance had not previously been named. The report notes the decision stems from ByteDance’s particular circumstances — the TikTok ownership question makes its relationship with the U.S. government especially sensitive, and any ammunition construed as “copying a U.S. model” would cost more than technical reputation. In other words, this is both a technology-roadmap choice and a compliance defense.

[HIRING & DELIVERY SHIFT] The delivery cadence of ByteDance’s AI team will be slowed by this discipline, and its near-term standing on domestic leaderboards will be hard to keep respectable. The pressure falls on team leads’ resource requests: skipping shortcuts means paying for more self-training compute and longer iteration cycles, all of which must be carved out of the budget in advance. It is also a variable on the hiring side — researchers willing to wait out long cycles and engineers eager to climb leaderboards quickly are not the same people you recruit.

▪ SIGNAL Under compliance pressure, technical purism is sometimes the most cost-effective insurance policy.

❯ Wayve Wins London Ride-Hailing License, Will Launch Safety-Supervised Autonomous Rides with Uber

[LICENSE SECURED] Transport for London has issued ride-hailing licenses to 15 electric vehicles from autonomous-driving company Wayve. The cars are Ford Mustang Mach-Es fitted with Wayve’s AI Driver system, surround-view cameras, and radar. Uber and Wayve confirmed they will first run safety-supervised passenger service before any full rollout — the vehicle drives itself while a licensed driver stays on board throughout, ready to take over at any moment. Fully driverless rides are not yet permitted.

[AHEAD OF RIVALS] The key to winning this license was assembling all of London ride-hailing’s “three-lock” requirements: operator, driver, and vehicle must each hold a license issued by the same regulator — not one can be missing. No autonomous-driving company had previously held all three at once. Market demand was already waiting — more than 100,000 Londoners joined Uber’s waitlist over the past eight weeks, queuing for the first rides. The competitive landscape has been redrawn: Waymo plans to launch in London in 2026, Baidu is also pushing ahead, but Wayve has put cars on the street first — safety supervisor on board.

[REGULATORY PATH PAVED] UK regulators now have a citable precedent. Latecomers will no longer be negotiating over “can we hit the road” — the question is which documents they need to complete under this licensing structure. The first one squeezed out of the time window is Waymo’s London plan — by the time it lands in 2026, local rivals will already have logged a full cycle of supervised, real-world mileage.

▪ signal: The first door for autonomous driving in London was opened with a ride-hailing license.

OUTLOOK

[TODAY'S BATCH] These four stories are all one story: AI’s cost structure is being rewritten in a scramble from every side. Anthropic is moving into chips, Meta has slashed coding-model prices to below DeepSeek’s, Musk has locked down next year’s GPU supply, and CXMT for the first time dared to hold its ground on price against Apple — every move lands on “who gets to set the price.” The Microsoft document supplied the answer from the opposite direction: when a single customer accounts for 70% of your AI revenue, pricing power sits with the other side.