Anthropic Hits $380B and OpenAI Sidesteps Nvidia [Model Behavior]
[00:00] Chad Thompson: Welcome to Model Behavior.
[00:03] Chad Thompson: Model Behavior examines how AI systems are built, deployed, and operated in real professional environments.
[00:11] Chad Thompson: Joining us today is Chad Thompson, who brings a systems-level perspective on AI automation and security.
[00:19] Chad Thompson: Chad, it is great to have you.
[00:21] Nina Park: We have a significant volume of industrial news to cover this morning,
[00:25] Nina Park: specifically regarding the capitalization of the major labs
[00:29] Nina Park: and a notable shift in the hardware ecosystem.
[00:32] Nina Park: Chad, we're seeing Anthropic reach a valuation of $380 billion
[00:37] Nina Park: following a $30 billion funding round.
[00:40] Nina Park: What does this scale of investment tell you about the current enterprise landscape?
[00:44] Thatcher Collins: It suggests the market is pricing in deep vertical integration.
[00:48] Thatcher Collins: Anthropics revenue is now $14 billion growing tenfold annually.
[00:53] Thatcher Collins: Their partnership with Infosys is the real story here.
[00:57] Thatcher Collins: They aren't just selling chat.
[00:59] Thatcher Collins: They are building agents for regulated industries like telecom and finance.
[01:04] Thatcher Collins: This requires a level of governance and reliability that goes beyond simple demos.
[01:10] Chad Thompson: Right. That drive for reliability seems to be pushing labs toward more specialized hardware.
[01:18] Chad Thompson: OpenAI just released GBT-5, 3B Codex Spark.
[01:23] Chad Thompson: And for the first time, they are running a production model on non-NVIDIA hardware.
[01:28] Nina Park: That's correct, Nina.
[01:30] Nina Park: They're utilizing Cerebrus Waferscale chips to achieve throughput of 3,000 tokens per second
[01:39] Nina Park: on their open-weight GPT-120B model.
[01:44] Nina Park: This shift towards Cerebrus is specifically for inference speed encoding tasks.
[01:50] Nina Park: OpenAI has reportedly been looking for alternatives to NVIDIA to reduce latency in real-time collaboration environments.
[01:59] Thatcher Collins: The speed is critical for the computer use features we're seeing.
[02:03] Thatcher Collins: Anthropic released Claude's Sonnet 4.6 yesterday, which is now the default for pro users.
[02:10] Thatcher Collins: It handles multi-step actions like filling out web forms and coordinating data across browser tabs.
[02:16] Thatcher Collins: When you combine that with the open-source growth of projects like OpenClaw,
[02:20] Thatcher Collins: which now supports these 4.6 models, the automation layer is becoming very dense.
[02:26] Chad Thompson: While the enterprise side focuses on automation, the consumer space is seeing its own consolidation.
[02:33] Chad Thompson: Apple is preparing to showcase Google Gemini integration in Siri later this month.
[02:38] Chad Thompson: Thatcher, this seems like a tactical move to address the capabilities gap Siri has faced for several years.
[02:45] Nina Park: It is a pivotal bridge.
[02:48] Nina Park: This late February rollout is the first public result of the Apple-Google partnership.
[02:54] Nina Park: It focuses on immediate utility while a more extensive Siri transformation is planned for later this year.
[03:02] Nina Park: Meanwhile, Google is already setting the stage for its next cycle, announcing Google I.O.
[03:09] Nina Park: 2026 for May 19th, where we expect further Gemini integrations into Android and Chrome.
[03:16] Nina Park: We should also note the shift in model reasoning.
[03:20] Thatcher Collins: GPT-52 Pro recently solved an Erdis conjecture in combinatorics.
[03:27] Thatcher Collins: Terence Tau verified the proof as quite different from previous human attempts.
[03:30] Thatcher Collins: It's the third Erdus problem resolved with AI assistance, following successes by the Aristotle system.
[03:39] Thatcher Collins: We're moving from models that retrieve information to models that can genuinely contribute to open mathematical problems.
[03:45] Chad Thompson: That's notable.
[03:47] Chad Thompson: It's a clear trajectory from high-level funding to specialized hardware and verified reasoning.
[03:53] Chad Thompson: We'll be watching how the market reacts to Alphabet Shares, which fell over 4% recently due to these rising capital expenditures.
[04:01] Chad Thompson: Chad, thank you for your insight today.
[04:03] Nina Park: It has been a significant day for the sector.
[04:07] Nina Park: Thank you for joining us.
[04:09] Chad Thompson: Thank you for listening to Model Behavior, a Neural Newscast editorial segment.
[04:14] Chad Thompson: Visit mb.neuralnewscast.com.
[04:19] Chad Thompson: Neural Newscast is AI-assisted, human-reviewed.
[04:24] Chad Thompson: View our AI transparency policy at neuralnewscast.com.
