Jeff Dean Exits Google, Hassabis Steps Back: AI News Aug 6

I. Today's Headlines

Google rebuilt the command structure of its AI division in a single announcement, and lost four of its most decorated engineers doing it. On Wednesday, August 5, Alphabet said Demis Hassabis will step down as chief executive of Google DeepMind and become its chairman, while taking a newly created title of Alphabet chief scientist. In a memo to staff he framed the move around the approach of artificial general intelligence: he wants time to work on research and strategy tied to the societal impact of AGI, will keep only a handful of direct reports, and will spend more of his week at Isomorphic Labs, the DeepMind drug-discovery spin-out he founded and still runs. "I've always believed the No. 1 application of AI should be to improve human health," he wrote. "It's time for AI to prove its unequivocal value to the world, and what better way to demonstrate that than to help finally cure diseases like cancer." Sundar Pichai's companywide note said the two had been discussing the arrangement for some time: "Demis and I have been long discussing a role that allows him to put his full attention on actively shaping the future of AGI." (Sources: Reuters, August 5; Alphabet internal memos as reported)

The operational job goes to Koray Kavukcuoglu, until now Google DeepMind's chief technology officer and Alphabet's chief AI architect. He takes the title of senior vice president rather than CEO, reports directly to Pichai, and owns Gemini model development, frontier research and the developer teams. A thirteen-year DeepMind veteran who worked on WaveNet and DQN, he is now the named owner of Gemini 4, the next flagship, which is still unreleased. No new CEO has been appointed. (Sources: Reuters, August 5; BigGo Finance, August 6)

The departures are the part the industry is reading hardest. Jeff Dean — Google's roughly thirtieth employee, hired in 1999, co-founder of Google Brain in 2011, a driving force behind the TPU program, and chief scientist since the 2023 Brain–DeepMind merger — is leaving after 27 years. He goes with Sanjay Ghemawat, his longtime collaborator on MapReduce and Bigtable, plus Oriol Vinyals, a key contributor to AlphaFold, and Quoc Le. The four are founding Discovery Loop, a public benefit corporation with Dean as chief executive, aimed at automating the multi-step loop of scientific and engineering work — propose a hypothesis, design an experiment, run it, evaluate the result — starting with machine-learning research and engineering itself before moving into hardware design, drug discovery and clean energy. "Many scientific research processes involve repeated steps that can be fully computerised," Dean said, explaining the company's name. Ghemawat was blunter about why they left: "We wanted to build something differently than how things are built at Google right now," noting Google's framework is optimised for massive consumer apps, search and advertising rather than specialised research infrastructure. Per the Wall Street Journal, Alphabet is a founding investor and cloud partner in the startup, with Radical Ventures and Khosla Ventures co-leading the seed round; Dean reportedly began planning it only about five weeks ago. (Sources: Reuters, August 5; Wall Street Journal via BigGo Finance; Dean's public post)

Markets read it as brain drain rather than succession planning. Alphabet fell as much as 5.4 percent intraday. The context is unforgiving: the flagship version of Google's next Gemini model missed a planned June launch and remains unshipped, Noam Shazeer left for OpenAI in June less than two years after Google paid billions to bring him back, and at May's developer conference Pichai and his lieutenants pitched Gemini on cost advantage rather than state-of-the-art capability. Alphabet also disclosed its first negative free cash flow since it began reporting the figure, with full-year capital expenditure guided to the top of a range near 205 billion dollars. Alphabet gave no explanation for why all of these moves landed on the same day. (Sources: Reuters, August 5; Cailian Press, August 6; BigGo Finance)

II. Model Releases and Product Updates

ByteDance put a native audio-video full-duplex model into every copy of Doubao. On August 5, ByteDance Seed released SeedRealtime, a single end-to-end architecture that natively fuses audio, video and text so the model perceives, understands, decides and speaks over a continuous multimodal stream — "watch, listen and talk at the same time." Unlike cascaded stacks that chain ASR, a vision model and TTS, there is no external voice-activity detector deciding whose turn it is, which removes both the latency and the information loss of module hand-offs. ByteDance claims three capabilities: joint audio-video understanding (using the picture to disambiguate homophones, resolve gestures and referring expressions, and track scene changes over time), proactive interaction (flagging when a target object appears, calling tools mid-conversation), and conversational rhythm control (knowing when to speak, pause or stay silent, and distinguishing the user from bystanders and background noise). Its end-to-end human evaluation reports that rhythm failures — being cut off mid-sentence, responding too slowly, or false-triggering on background chatter — fell by roughly half versus a cascaded system, with a higher probability that a single conversation runs to completion smoothly. Demos included identifying who is speaking at a multi-person gathering, reading a Chinese menu for a foreign tourist in real time, watching a museum camera feed and speaking up when a specified artefact appears, and correcting a user's coffee-machine operation as the picture changes. It is fully rolled out in the Doubao app under the "call" option in the chat box. (Source: ByteDance Seed, August 5)

Nvidia open-sourced two foundation models for physical AI. Alpamayo 2 Super, released for commercial use, is an open reasoning model for robotaxis and autonomous vehicles: 34 billion parameters pairing a 32B Cosmos 3 Super Reasoner with a 2B diffusion-based Action Expert, producing driving trajectories alongside chain-of-causation reasoning traces, predicting meta-actions such as yielding or lane changes, and answering natural-language questions about a driving scene with 2D visual grounding. Nvidia reports it ranks first among nearly forty models on autonomous-driving benchmarks while remaining deployable for on-vehicle inference. Alongside it, Cosmos 3 is an open foundation model for World Action Models, which predict how the physical world evolves rather than merely describing a scene — a Mixture-of-Transformers trained on 767 million images, 348 million videos and 8 million action samples, shipped with trained policies including Cosmos3-Nano-Policy-DROID. Nvidia's argument is that a dynamics-aware model lets robotics policies generalise to new environments and embodiments from fewer demonstrations. (Source: Nvidia, August 5)

Tencent Hunyuan shipped a new speech recogniser; xAI's voice alias flipped over. Tencent Hunyuan released Hy ASR 3.0 preview on August 4, its next-generation speech recognition model. Separately, xAI's grok-voice-latest API alias migrated on August 5 to Grok Voice Think Fast 2.0, the speech-to-speech model xAI announced on July 29 with improvements to intelligence, transcription accuracy and tool-use reliability — the announcement is old news, the default routing change is not. (Sources: Tencent Hunyuan, August 4; x.ai, July 29)

Microsoft is testing a full-duplex voice model of its own. A hidden entry in Microsoft's MAI Playground revealed MAI-Realtime, a first-party bidirectional voice model with two voices, Victoria and Grant, and support for 17 languages, positioned directly against OpenAI's GPT-Live. Limited partner access is reportedly underway; no public launch date has been announced. (Source: TestingCatalog; subject to vendor confirmation)

OpenAI published two updates on August 4. One covers third-party cyber evaluations involving OpenAI models, the other introduces new ways to learn and teach with ChatGPT Work and Codex. (Source: OpenAI, August 4)

III. Industry and Capital

Anthropic signed a 10 billion dollar, six-year compute contract for a single Norwegian site. Volta Infra Holdings announced on Tuesday that it had landed a six-year, 10 billion dollar agreement with an unnamed frontier lab, to be delivered with Bitdeer Technologies Group at a data centre in Norway; Bloomberg identified the customer as Anthropic, citing people familiar with the matter. Anthropic declined to comment and Volta did not respond to a request for comment. The site offers roughly 133 gross megawatts and 121 critical IT megawatts, to be fitted out entirely with Nvidia's next-generation Vera Rubin systems. Bitdeer's own disclosure describes a sixteen-year lease and services agreement carrying about 4.7 billion dollars in scheduled payments over the initial term at an average of roughly 202 dollars per kilowatt-month with 3 percent annual escalators, backed by about 1.3 billion dollars of credit support in letters of credit arranged by affiliates of J.P. Morgan and another institution. Bitdeer puts remaining capex at about 500 million dollars and will hand over capacity in four data halls in two phases, targeting December 31, 2026 and March 31, 2027. Volta was founded in January by former Brookfield Asset Management executives, has raised 300 million dollars at a 2.4 billion dollar valuation from backers including Nvidia, Andreessen Horowitz, Altimeter and Azora, and separately announced a 5 billion dollar financing programme with Azora. Bitdeer, a Bitcoin miner redirecting facilities toward AI, rose nearly 10 percent premarket. The deal joins Anthropic's existing arrangements with SpaceX, AMD, Akamai, Amazon, Google and Broadcom, and Microsoft and Nvidia; it is also reportedly in talks to lease capacity from Meta's data centres. (Sources: Bloomberg via Yahoo Finance and Economic Times, August 4; Bitdeer press release; Reuters)

Meta started renting out its own GPUs. On August 5 Meta said it is commercialising its compute, selling capacity to external customers — a reversal that turns one of Nvidia's four largest buyers into a competitor in the infrastructure market. Microsoft, Google, Amazon and Meta together accounted for roughly half of Nvidia's data centre revenue over the past three years; Nvidia's fiscal 2026 revenue was 215.9 billion dollars, of which data centre contributed 193.7 billion, close to 90 percent. All four are simultaneously scaling in-house silicon — Google's TPU at volume, Amazon's Trainium accelerating into training, Microsoft's Maia 100 in production inference, Meta's MTIA in recommendation systems — while OpenAI develops an inference chip and Apple works with Broadcom. The substitution is concentrated on inference, where per-card efficiency, power and unit cost dominate; training still favours Nvidia's compute density, interconnect bandwidth and software maturity. (Sources: Chinese tech press, August 5; subject to vendor confirmation)

AI earnings kept clearing the bar, and the market noticed. Amazon became the fifth US technology company to pass 3 trillion dollars in market value on August 4, with AWS posting its fastest growth in eighteen quarters. Palantir reported second-quarter revenue up 93 percent year on year and raised full-year guidance to 8.15 billion dollars; the stock rose 29.5 percent, its largest single-day gain since February 2024. AMD reported second-quarter revenue of 11.54 billion dollars, up 50 percent and ahead of an 11.31 billion estimate, with adjusted earnings per share of 1.66 dollars, up 246 percent. The Dow and S&P 500 both closed at record highs and the Philadelphia Semiconductor Index gained more than 7 percent. Castle Securities projected that US technology companies will issue more than 500 billion dollars in debt to fund AI chip purchases through 2028, potentially reshaping the investment-grade credit market. (Sources: Yicai, 21st Century Business Herald, Cailian Press, August 4–5)

Washington moved on Chinese data-centre hardware. Reuters reported on August 4 that the US government is drafting a ban on imports of new Chinese-model data centre components, with the Federal Communications Commission preparing measures targeting optical transceiver modules used for high-speed fibre transmission inside data centres. Officials want it published and in force within 2026, though people familiar said the FCC could still amend or shelve the restrictions. Coherent and Marvell rose more than 12 percent, Corning more than 9 percent. Separately, Anthropic named Mariano-Florentino "Tino" Cuéllar chief global affairs officer on August 4, and SK Hynix and SanDisk published the first global standard specification for High Bandwidth Flash, with Google joining the HBF alliance. Nvidia open-sourced its cuFile APIs and launched Storage-Next with more than forty storage vendors, with Google, Intel and Meta as inaugural maintainers. (Sources: Reuters, August 4; Anthropic, August 4; Cailian Press; Nvidia, August 5)

IV. China's AI Landscape

DeepSeek restarted a funding round that had been abruptly paused. Caijing reported on August 5, citing multiple people involved in the transaction, that DeepSeek has resumed its second funding round, seeking 50 billion yuan at a pre-money valuation of about 500 billion yuan, with signing expected in late August. The round originally launched in mid-July and was suspended at the end of July, when investors on the waiting list were told signing plans were on hold; reporting has attributed the pause in part to founder Liang Wenfeng's displeasure at widely circulated commentary around an apparently leaked investor meeting transcript. DeepSeek has not responded to the report. Its first round opened in April and closed in June at 50 billion yuan on a valuation above 350 billion yuan, the largest first round in the history of Chinese large-model companies. (Source: Caijing via 163 Tech, August 5; subject to company confirmation)

ByteDance is spending to match. Beyond SeedRealtime, ByteDance has reportedly raised its 2026 AI infrastructure capital expenditure budget by about 25 percent to 200 billion yuan, of which roughly 85 billion is earmarked for AI chips, with memory chip costs cited alongside rising AI investment as the driver. On August 3 the company opened its class-of-2027 campus recruiting with new "AI full-stack engineer" and "AI agent developer" roles appearing for the first time. Its Seedance 2.5 video model, released July 31, extended single-shot generation from 15 to 30 seconds and accepts up to 30 images, 10 videos and 10 audio clips as references in one prompt. (Sources: Chinese tech press, August 3–5; ByteDance Seed, July 31; capex figures subject to company confirmation)

Chinese open-weight models still hold the OpenRouter usage crown. On the aggregator's most recent weekly token-call leaderboard, the top five entries were all developed by Chinese companies. Xiaomi's MiMo-V2.5 led with 10.5 trillion tokens in a week, up 12 percent, having gone to public beta on April 23 and fully open-sourced by the end of April. DeepSeek held second and fifth. Tencent Hunyuan's Hy3, open-sourced on July 6, took third with week-on-week growth above 999 percent. OpenRouter aggregates paid calls from tens of thousands of independent developers and small AI applications outside China on a single token accounting standard, though it excludes private enterprise deployments and closed US API traffic. (Source: CCTV via Toutiao, August 2)

Regulatory watch: export controls may run in reverse. China's Ministry of Commerce has reportedly been consulting Alibaba, ByteDance and Zhipu on export controls that would invert the current playbook — restricting Chinese firms from sending training data overseas, limiting foreign access to Chinese open-weight model downloads, and potentially barring Qualcomm and TSMC from manufacturing chips built on Chinese firms' own designs. No final rule has been issued. Separately, Baidu consolidated its AI office products, folding dodo into Baidu Dazi. (Sources: The Leading Edge newsletter, week of August 3; Chinese tech press; subject to official confirmation)

V. Today's Observation

Yesterday's story was about who is allowed to test frontier models. Today's is about who is left to build them.

The Google reshuffle is easy to misread as an ordinary succession. It is not. Hassabis did not get promoted out of DeepMind so much as he traded operating authority for research latitude at the exact moment the lab most needs an operator — Gemini 4 is late, the flagship slipped a June date, and the person now accountable for shipping it holds an SVP title rather than a CEO one, reporting into Pichai. That is a deliberate flattening. Alphabet has decided the bottleneck is execution, not vision, and has restructured to put the model roadmap one reporting line from the top.

The Dean departure cuts differently. Losing one star researcher to a rival is a compensation problem. Losing four founding-generation engineers as a unit, to a company they built in five weeks, is an infrastructure problem — and Ghemawat said so out loud: Google's stack is tuned for consumer scale, not for specialised research. That Alphabet then chose to invest in Discovery Loop and serve as its cloud partner is the most telling detail of the day. It is a hedge dressed as generosity: keep the optionality on whatever Dean builds, without carrying the organisational cost of building it in-house. Expect more of this shape. The frontier labs are large enough that their best research now has a lower cost of capital outside the company than inside it.

Meanwhile the compute story keeps hardening into something that looks less like procurement and more like project finance. Anthropic's Norwegian contract is not a cloud invoice; it is a six-year obligation attached to a specific power block, a specific chip generation and dated construction milestones, underwritten by letters of credit and a sixteen-year lease on the other side. Stack it against the SpaceX, AMD, Amazon, Google-Broadcom and Microsoft-Nvidia agreements and the shape of the risk becomes clear: a frontier lab's balance sheet is now a portfolio of long-dated, illiquid capacity commitments priced off demand forecasts that nobody can validate. Meta renting out its own GPUs on the same day is the flip side of the same trade — if you have overbuilt, the capacity itself becomes the product. Both moves make sense individually. Together they describe an industry that has moved its bets from models to megawatts, where a miss shows up not as a bad benchmark but as a stranded asset.

---

Compiled from vendor announcements and public reporting on August 5–6, 2026. Aggregated and media-sourced items are subject to vendor confirmation.