Weekly Digest 32

Topics we are tracking

Google DeepMind loses its CEO and its chief scientist on the same day

Source: Axios — Google DeepMind CEO Demis Hassabis is stepping aside

Demis Hassabis is leaving his role as CEO of Google DeepMind to become chairman of the unit, adding the title of chief scientist at Alphabet while continuing to run Isomorphic Labs. On the same day, chief scientist and Google legend Jeff Dean is leaving the company alongside several Google AI colleagues to start a company of their own. Koray Kavukcuoglu, until now chief technology officer of Google DeepMind, becomes senior vice president of the unit and reports directly to Sundar Pichai.

We have previously written about the departures of Noam Shazeer and John Jumper and dismissed them at the time as minor. This reorganisation is very different, and it is more of a warning sign. Hassabis has widely been believed to be the successor to Sundar Pichai as CEO of Alphabet. Perhaps one of the finest scientific minds of our generation, he has been vital to projects like AlphaGo and AlphaFold, the latter of which won him the Nobel Prize. DeepMind has pursued many incredible scientific endeavours, much in line with Hassabis's own philosophy, and has failed to monetise them so far.

Dean is widely regarded as one of the most important figures in the company's history and was employee number 30. He led the creation of Google Brain, which was later folded back into Google DeepMind.

His new company, Discovery Loop, which he founded together with the other longtime Google veterans Quoc Le, Oriol Vinyals and Sanjay Ghemawat, is focusing on recursive self-improvement: automating the machine-learning research loop itself, with hardware design and drug discovery as later targets. Discovery Loop is set up as an independent public benefit corporation, with Google as an investor and its cloud provider.

This marks the second high-profile neo-lab explicitly built around recursive self-improvement. Recursive Superintelligence was founded only months ago, in late 2025, and emerged from stealth in May with $650 million at a $4.65 billion valuation from GV and Greycroft, with Nvidia and AMD participating. There seems to be real momentum behind recursive self-improvement at the frontier labs. More than 1,100 employees of OpenAI, Anthropic, DeepMind and Meta, including Dario Amodei and OpenAI's chief scientist, signed the “Pacing the Frontier” petition in late July, asking the US government to build the tools to deliberately pace automated AI research. Nobody petitions for a brake on something they consider speculative.

Everything here points in the same direction. Google bureaucracy has increasingly been a frustration for many researchers, Gemini 3.5 Pro is reportedly months behind schedule, and the company has made a conscious choice to sell TPU capacity to Anthropic rather than keep it for internal use, leaving research teams short of compute. Even the Discovery Loop deal carries the same signature. Google would rather be the investor and the compute vendor to its best researchers than the place where they do the work.

The reporting line tells the same story. DeepMind is no longer run by a scientist-CEO with a mandate of his own, it is run by a senior vice president inside the corporate structure. This all suggests that Google will move closer to Microsoft and Amazon and further away from OpenAI and Anthropic. Both keep internal research teams working on their own models, neither of them anywhere near the frontier, and both are far more focused on building out compute and becoming a hub for a variety of different models. At this point you cannot call Google a frontier lab.

Google has all the ingredients of being one, perhaps more than anyone in the modalities outside of language. But it got lost in the weeds, with too many pursuits and not enough compute to double down on any of them. Perhaps Sergey Brin taking a more active leadership role again will help the company make the hard decisions and cut the projects that are pipedreams and going nowhere.

SK Hynix and SanDisk publish the first High Bandwidth Flash standard

Source: TrendForce — SK hynix, SanDisk Debut HBF Standard to Challenge AI Memory Bottlenecks with Google, Tenstorrent Support

SK Hynix and SanDisk have published the first standard specification for High Bandwidth Flash. HBF stacks NAND the way HBM stacks DRAM, creating a new memory tier between HBM and SSDs. The specification supports up to 512GB per device in 8-high and 16-high stacks and three bandwidth grades running from roughly 0.4 to 3.0 TB/s, and it uses the UCIe interconnect, so it attaches to GPUs and CPUs alike. It was announced at the Future of Memory and Storage conference and published through the Open Compute Project, with Google DeepMind and Tenstorrent backing the consortium.

AI inference is increasingly constrained by memory capacity per accelerator, particularly on long-context and agentic workloads. Model weights and the KV cache need to sit somewhere they can be reached quickly. HBM has the bandwidth but is limited in capacity and extremely expensive, while SSDs have the capacity and are far too slow. HBF sits between the two and could significantly reduce the cost of inference.

SanDisk plans first HBF prototypes in the second half of this year, a pilot line in Japan should begin operating around year end, and commercialisation is targeted for 2027. SK Hynix has also announced its 375-layer V10 4D NAND, with mass production in early 2027 and 2.5× better performance per watt, and now claims the industry's only full-stack AI memory portfolio spanning HBM, HBF and eSSD.

TrendForce expects demand for complex memory solutions including HBF to pick up only around 2030. There are also good reasons for scepticism. NAND endurance and write latency make HBF a read-mostly tier for weights and cache rather than a DRAM replacement, and previous attempts to insert a new layer into the memory hierarchy, Optane being the most famous, all died.

The technical case for HBF in inference is nonetheless strong. Inference has a largely deterministic memory access pattern. Weights need no random access at all, and a cold start of a few milliseconds is tolerable as long as the continuous reads that follow arrive at full bandwidth. That is precisely what NAND is good at, and the cost per bit is an order of magnitude below HBM even after the stacking and logic overhead.

Whether that is enough to overcome the software work remains to be seen. Nvidia, so far at least, does not seem interested. Some observers have speculated that HBF will appear in Nvidia products from 2027 onwards, but there is no concrete evidence for that.

ByteDance is reportedly training a 10-trillion-parameter model

Source: Reuters — ByteDance targets mega AI model nearing Anthropic's Mythos, FT reports

ByteDance is training an AI model with up to 10 trillion parameters, according to an FT report citing people with knowledge of the matter. That would be more than three times the size of Moonshot's Kimi K3 (2.8 trillion, currently China's largest released model) and close to industry estimates for Anthropic's Mythos 5 (around 8 trillion, with Fable 5 estimated at 5 trillion). The model is in pre-training, which typically takes three to six months, and the final size has not been decided.

After focusing on efficiency through 2025, Chinese labs have now fully embraced the Bitter Lesson and shifted more towards scale. Following the Kimi K3 release, Meituan and Alibaba also put out trillion-parameter models, all of them larger than the largest American open-weight models. There is no information yet on whether this model would be open weights, but we would be surprised if it were closed, given Xi Jinping's recent commitments to the open ecosystem.

Chinese companies have historically allocated more of their compute to non-LLM businesses such as recommender algorithms, model inference and renting out capacity through cloud services. That appears to have changed in recent months, with more compute moving toward pre-training. ByteDance owns the most popular AI consumer app, Doubao, with 324 million monthly active users, and one of the world's most advanced video models in Seedance. Serving those users already takes a large GPU fleet, and a 10-trillion-parameter pre-training run would absorb a large share of ByteDance's overall capacity.

Some back-of-the-envelope math: A Mythos-level model was trained on somewhere around 330,000 H100 equivalents running for about three months. ChinaTalk estimates China holds 2.8 million H100 equivalents, including domestic chips, remote access, smuggled and legally purchased units.

ByteDance is reportedly Oracle's largest GPU customer. Its demand helped make Johor, Malaysia the world's second-largest AI hub, and it is now adding about 36,000 B200 GPUs through Aolani Cloud with Nvidia's support. The export-legal structure means that offshore capacity can only be used for work conducted outside China. How well that is enforced is not clear to us and needs further investigation.

Seen on X

Other interesting stories

• Bloomberg: Citadel Securities Says Bull Market Drivers Are ‘Firmly Intact’
• The New York Times: A.I. Designed These Viruses From Scratch
• AI Futures Project: How to pace the US frontier
• Reuters: Uber forecasts weak quarterly profit, doubles down on robotaxi investment plans
• The Verge: Texas orders an audit of data centers seeking grid connections