Latest / UpNext AI / Anthropic and Nscale’s Compute Talks, ProteinGeoBench, and LG-Bench
Transcript
- Welcome to the UpNext AI podcast. It's Tuesday, October 20th, 2026, and here's what matters in AI today. Anthropic and infrastructure provider Nscale are discussing a potential compute deal worth $45 billion, according to Bloomberg reporting cited by TechCrunch. The talks are ongoing, and the terms could change. The companies announced a separate compute deal worth $45 billion a little over a month ago. This would be another agreement, rather than a new price tag for that earlier one. The scale matters because access to computing capacity shapes how frontier labs develop and operate their systems. Anthropic has struck a series of large compute deals over the past year, and another agreement with Nscale would add to those commitments. A proposed dollar amount, though, does not tell customers when capacity will arrive or how an agreement will be financed. Those questions remain separate from the talks themselves. There is another Nscale development worth keeping alongside this one. The Financial Times reports that Microsoft is calling time on its Nscale deal. That is a separate partnership, and its change does not establish what will happen in Anthropic's negotiations. For anyone assessing AI infrastructure plans, the immediate takeaway is to distinguish a reported negotiation, an announced agreement and delivered capacity. That question of who can access AI capability extends beyond data centers. A piece titled “Agentic inequality” in Nature Machine Intelligence brings the distribution of AI agents into view. Its cited work spans the digital divide, computing access, labor-market effects and governance. Those are useful lenses for organizations deciding how agents should be made available and who controls the systems they depend on. The piece is a prompt to examine access and oversight; the citation list does not establish a measured effect of agent deployment. Access is only part of the story. Capability depends on the task you ask an agent to perform. ProteinGeoBench tests that distinction in structural biology: an agent may retrieve recorded information about a protein while struggling to calculate spatial relationships from its coordinates. In the preprint posted Monday, researchers evaluated protein-structure tasks and found that three of the four frontier models tested did better on metadata lookup than on numeric calculation. The researchers also found that evaluation choices matter. With a single run per task, repeated statistical samples sometimes named a different top model. Changes to the agent setup also shifted scores, although that comparison was too small to settle the size of the effect. If you are evaluating a scientific agent, ask for results by task type and for the access conditions and number of runs. A lookup score alone cannot establish reliable calculation, and these benchmark tasks do not establish performance across every scientific workflow. A second life-sciences preprint asks a broader evaluation question: can a model answer related questions consistently, rather than merely score well across a disconnected set? LG-Bench links more than 10,000 multiple-choice questions in medicine, biology and chemistry according to their semantic relationships. The researchers propose measures for consistency among related questions and for where errors cluster, so evaluators can look for conceptual blind spots. That design could help teams inspect a model's pattern of mistakes instead of relying on one overall accuracy figure. The paper describes the benchmark and its proposed measures; its abstract does not establish that a particular model performs better or worse under them. For now, the useful takeaway is an evaluation method: related questions may reveal weaknesses that a single aggregate score obscures. ...Are you building apps with voice? Elevate your app's voice capabilities with ElevenLabs. Their API is a game changer for embedding dynamic, responsive voice interactions in your applications, providing unprecedented realism, flexibility and latency. Visit up next dot fm slash eleven to check out their latest offerings. ... A couple of other developments before we go. TechCrunch reports that Nvidia challenger Tenstorrent raised $1.5 billion at a valuation of $6.2 billion after drawing acquisition interest earlier this year. The funding supports its decision to remain independent as it pursues open-weight AI chips. And AI assistant startup Town announced a Series B round of $108 million, co-led by Forerunner Ventures and Menlo Ventures. TechCrunch reports that the company reached a unicorn valuation four months after launch. Before we wrap up, a quick note: this podcast is generated with the assistance of AI and is intended for informational purposes only. All referenced articles, research, and commentary remain the property of their original authors and publishers. And that's your briefing for today. If you enjoyed this episode, share it with a friend or colleague—and follow UpNext AI on Apple Podcasts, Spotify, or wherever you get your podcasts! Full source links are in the episode notes, and we'll be back tomorrow with what's up next!