Morning briefing · 08:00
AI valuations and GPUs get repriced
The prices the AI industry put on capital and compute are being readjusted, and efficiency, doing more with fewer resources, is gaining weight as a result. On the same day, reports said Anthropic's expected IPO valuation had fallen and that Microsoft is winding down leased capacity based on older GPUs. Alongside this, research and open source produced a string of techniques for running large models on limited hardware.
Price tags on capital and hardware are adjusting
The most direct signal came from capital markets. According to AI Times, citing a Bloomberg report, investors considering Anthropic's IPO value the company at $1.2–1.3 trillion, with some going as low as $1.1 trillion1. That is about 45% below the up to $2 trillion once floated. The recent sharp drop in AI stocks is cited as the reason, but the final offering price has not been set and next week's roadshow will go ahead as planned.
A similar reassessment was reported on the hardware side. According to Bloomberg, citing multiple sources, Microsoft has reduced the compute it leased from Oracle over several months and decided not to renew existing contracts2. The capacity was rented in 2023 when infrastructure was scarce, but older NVIDIA GPUs no longer meet current requirements, according to the report. Oracle shares fell as much as about 3% intraday after the report.
The two stories cover different areas but point the same way: values set during a scramble for supply a few years ago are being recalculated based on present-day usefulness. In particular, it has become clear that GPU capacity is an asset whose value can drop quickly when generations change.
How efficiency extends the usefulness of hardware
While older GPUs in data centers are being repriced, the software layer saw continued efforts to extract more performance from existing hardware. Tencent researchers published a paper on 'DSpine', a drafter for speculative decoding (a technique that speeds up inference by having a small drafter propose several tokens first and a large model verify them all at once)4. It targets the problem of candidate tokens drawn in parallel becoming inconsistent and shortening accepted spans, and proposes Adjacent Causal Injection, which feeds information about the preceding token into later tokens at earlier layers.
The same trend shows up in local inference. PR #30966 'llama_moe_cache', opened in the llama.cpp repository, computes MoE experts not loaded in VRAM on the CPU and adds a tuning script6. The Reddit poster said it substantially speeds up offloading in low-VRAM environments, but whether it has been merged is not confirmed. China's OpenCSG also released 'Agentic-27B', built for tool selection and failure recovery, along with test results run on a single A800 GPU5. Commercial use, however, requires separate permission.
All three focus on getting more out of the same compute without scaling up. As hardware gets repriced, such techniques become a way to extend the life of existing resources.
Scaling as a different choice
Efficiency gaining weight does not mean scaling strategies are retreating. John Langford wrote on his blog that the roughly 50-person Microsoft Research AI Frontiers group was disbanded after being reorganized into Microsoft AI3. He explained that Microsoft AI's focus on acquiring, generating, and improving data made it hard to align with the group's research direction, and argued that efficiency gains through research can deliver significant value independent of scaling. This is one person's account, and the company's position is not given in the sources.
The case shows that inside a large organization, scale and efficiency sit as competing options for resources. The evidence in the previous sections is also not yet solid. The valuation figures are reported investor estimates, and the offering price has not been set. The efficiency techniques are likewise at the paper and unmerged-PR stage, and their effects have not been confirmed in real serving environments.
Companies and practitioners now need to weigh GPU generation and replacement cycles when contracting for infrastructure, and to review inference efficiency techniques suited to the resources they have. When adopting open models, they should also check commercial-use terms in advance. What to watch next: the offering price and demand set at Anthropic's roadshow next week, whether the llama.cpp PR is merged, and whether techniques like DSpine make it into serving frameworks.