This is the evening edition for October 29, 2026. The morning brought Google's model offensive and midday delivered earnings numbers from open-model and AI semiconductor players, but this evening is dominated by regulatory and political risk news. The U.S. Department of Justice has moved on a major NVIDIA acquisition, and Anthropic has put its tensions with the administration on paper ahead of its IPO. The community side, meanwhile, serves up a lineup of stories where hand-crafted ingenuity shines.
DOJ Opens Inquiry into NVIDIA's Hugging Face Acquisition — Probing the Competitive Impact of the $13 Billion Deal
According to a Reuters report, the U.S. Department of Justice (DOJ) has opened an inquiry into NVIDIA's acquisition agreement with AI model company Hugging Face. The deal is valued at approximately **$13 billion**. Multiple people familiar with the matter say the DOJ is examining how the transaction could affect competition in the AI market. At the same time, a long-running investigation into NVIDIA's conduct as a chipmaker is reportedly still underway.
NVIDIA has spent the past few years expanding its investments and acquisitions from its position of control over GPU supply. Rising scrutiny from competition authorities is a natural consequence, and the deal looks set to become a focal point in debates over concentration in the U.S. AI industry. For now, the matter remains at the inquiry stage; there is no confirmation of an escalation to a formal investigation or any move to block the transaction.
Anthropic Discloses “Tensions with the Administration” as a Risk Ahead of Its IPO
Anthropic — reported earlier this week to be two weeks from an IPO at a $2 trillion valuation — has revealed an unusual risk in the registration documents filed ahead of going public. The company warns investors that “**candid advocacy on AI safety could become a target for government retaliation**.” The disclosure states that the company's positions on AI oversight and statements from its leadership could expose it to serious levels of government pressure — an implicit allusion to its rift with Trump administration officials over safe development.
What makes this unusual is just how explicitly “conflict with the government” is spelled out in a corporate risk disclosure. A year's worth of skirmishes with the administration has now worked its way into the financial documents of a soon-to-be-public company — a measure of how tense the political climate around AI policy has become. Whether it affects the IPO valuation is a key point to watch.
Microsoft and Alphabet, a Study in Contrasts on AI Spending — Azure Up 48%, Search Ads at a Five-Quarter Low
Earnings released by the two companies on the same day illustrate the dividing line in how the market judges the AI investment bet.
Microsoft's cloud business is accelerating. Azure revenue grew **48% year over year to $33.2 billion** in the quarter ended September. Total revenue climbed 19% to $92.3 billion, beating analyst forecasts on both counts (44% Azure growth and 17% overall). The company has also managed to rein in capital spending — one answer to the market's nagging question of whether AI spending will ever pay for itself.
Alphabet, on the other hand, reported revenue excluding partner payments of $111.2 billion, falling short of market forecasts. Advertising revenue from Search and related businesses — roughly half the company's total and the funding source for its massive AI investments — also missed expectations, with growth slowing to its **slowest pace in five quarters**. Even though the headline “25% Q3 revenue growth” we covered at midday looks solid, the details leave lingering concerns about being caught by AI rivals. The diverging fortunes of these two companies, both spending heavily on AI, sent their stocks in opposite directions on the same day.
Palantir Defense Chief: “AI Advances Could Prompt Action on Taiwan”
A geopolitical remark also caught the eye. Mike Gallagher, head of Palantir's defense business (a former U.S. House member who led a committee that was hawkish on China), warned that advances in AI capability are making a takeover of Taiwan by President Xi Jinping more attractive. His argument: as Taiwan grows more critical to AI development, Beijing's incentives to act against the island are strengthening — particularly amid the U.S.-China technology rivalry. Taiwan is the chipmaking hub underpinning the global AI boom, and the remark is a telling sign of just how tightly AI and security are becoming intertwined.
Shrinking NVIDIA's 30B Speech LLM from 67.5GB to 10GB — WER of 5.16% with Zero Retraining
There's an interesting community follow-up to NVIDIA's transcription-focused model “Nemotron 3.5 Transcribe,” which we covered this morning. In an attempt posted to LocalLLaMA, the 30B speech LLM was **compressed from 67.5GB down to 10GB** through structural pruning. Better yet, it involved no retraining whatsoever, and transcription accuracy held at a word error rate (WER) of 5.16% (the original model scores 4.12%). The compressed version is available on Hugging Face as “nemotron-3.5-transcribe-mini-64e.”
A 67.5GB model is an awkward size for consumer GPUs, but at 10GB it should run on a single 16GB-class VRAM card. Taking a massive fine-tuned model, slimming it down, and making it portable — engineering that's characteristic of local LLM culture, applied almost immediately after a major vendor's release. A great example of the genre.
LIMe — Extending Attention to Cut Training Compute by 15%, Countering Representation Collapse
On the research side, “**LIMe (Layer-Integrated Memory)**” from T-Tech, presented at EMNLP 2026 currently underway in Hungary, is drawing attention. The authors formalize the problem this way: in a conventional Transformer, each layer depends only on the hidden states of the layer immediately before it, so all contextual information must pass through that bottleneck — creating pressure toward “representation collapse,” where tokens grow too similar to one another and distinctive features are lost.
LIMe is a lightweight extension to the attention mechanism. It takes advantage of the key-value buffers each layer already maintains and introduces learnable routing weights per attention head and per layer. As a result, each head can retrieve and combine representations not just from the preceding layer but from earlier layers as well — reusing already-computed information while keeping hidden state sizes unchanged. In experiments at around one billion parameters, it reached the same training loss as the LLaMA baseline with **15% less compute**, and it outperformed competing methods on language modeling and multi-step reasoning benchmarks. The gains were reportedly largest on synthetic arithmetic tasks and in deeper architectures. Replication at larger scale is still pending, but it's an intriguing result as an option for cutting training costs.
A $60 Million AI Supercomputer Hits eBay — 9,600 A100s Across 300 Racks, Free Shipping
Finally, an oddball item. An IT vendor has listed a decommissioned GPU cluster — the whole thing — on eBay. Asking price: **$60 million**. The 300 racks hold 9,600 NVIDIA A100 GPUs and 4,200 AMD EPYC processors, and shipping is free. It's not exactly something an individual can bid on, but the sight of data center assets swept aside by the generational transition hitting the secondary market is a window onto the flip side of the AI boom. If the used A100 supply grows, it could even be a tailwind for local LLM enthusiasts.