Anthropic’s TPU move and NVIDIA’s Starcloud

The Mixture of Experts panel discusses Anthropic's major investment in Google's TPUs for inference, debates the feasibility of insuring superintelligence, critiques OpenAI's approach to handling sensitive user conversations, and explores the futuristic concept of data centers in outer space.

October 31, 2025 6 min read

Anthropic’s TPU move and NVIDIA’s Starcloud — source: https://www.youtube.com/watch?v=KF3jIUBecFo

Anthropic’s Strategic Bet on Google’s TPUs

The panel dives into Anthropic’s announcement of its plan to dramatically expand its use of Google Cloud, committing to using up to one million TPUs (Tensor Processing Units). This move is seen as a strategic diversification of their compute resources beyond the highly sought-after, and often scarce, NVIDIA GPUs.

Key discussion points include:

Diversification and Cost-Efficiency: The primary driver is likely to secure compute capacity and find a more cost-effective solution for inference at scale. With the difficulty of acquiring GPUs, diversifying their hardware portfolio across multiple clouds (AWS, Google Cloud) is a smart strategy to de-risk their operations and manage costs.
Training vs. Inference Workloads: The experts speculate that this move reflects a shift in compute needs from being training-heavy to inference-heavy. They suggest Anthropic will likely continue to train its novel, cutting-edge models on NVIDIA’s platform to leverage the mature and highly optimized CUDA ecosystem they have already invested in heavily. The TPUs, which are ASICs designed specifically for AI workloads, will be used for running inference on older, more established models.
Technical Complexity: While strategically sound, operating across different chip architectures introduces significant operational complexity. Anthropic’s team will need to manage different software stacks and kernel optimizations, as they cannot rely solely on NVIDIA’s CUDA. However, for well-understood tensor operations in older models, other driver packages like Vulcan and OpenCL are becoming more competitive.
Philosophical Differences with OpenAI: The panel contrasts Anthropic’s approach with OpenAI’s. While there are rumors of OpenAI pursuing vertical integration by developing its own chip, Anthropic appears more focused on leveraging existing platforms to meet practical, enterprise-scale deployment needs. This highlights a potential difference in philosophy: Anthropic solving today’s problems versus OpenAI’s long-term pursuit of AGI at all costs.
Hardware Evolution: The move is compared to the hardware progression in Bitcoin mining, which evolved from CPUs to GPUs and finally to specialized ASICs. The AI industry is following a similar trajectory, moving towards specialized chips for greater efficiency on specific tasks.

The Skepticism Around “Underwriting Superintelligence”

The conversation shifts to a recent essay proposing that the insurance industry could play a key role in managing the risks of AGI through an “incentive flywheel”—where insurers drive the adoption of safety standards and audits. The panel, however, expresses deep skepticism about the practicality of this idea.

The main arguments against a third-party insurance model for AI include:

First-Party Responsibility: The model developers—multi-trillion-dollar companies like Google, Microsoft, and OpenAI—are best positioned to understand and underwrite the risks of their own models. They are already offering forms of first-party indemnification, such as for copyright infringement, because they have the deepest insight into how the models were trained and the data used.
Lack of Expertise and Transparency: Third-party insurers would lack the deep technical expertise to accurately assess the risk. The essay itself notes there are only about 100 researchers in the US qualified to conduct model audits. Furthermore, it’s highly unlikely that model developers would grant insurers the necessary access to their proprietary architectures and weights.
The “Lawyer Problem”: One panelist humorously notes that insurance inevitably leads to complex clauses, gray areas, and legal battles where the insurer will likely shift blame to the user for improper implementation (e.g., “you didn’t use the right safety rail”).
Risk is Use-Case Specific: A fundamental flaw in the concept is that AI risk is not general; it’s highly specific to the application. A general insurance policy for a model is naive. Regulation and risk management already exist for high-stakes domains like biomedical applications, and this is where the focus should remain.
Global Competition: Insurance is a regional market. Imposing such a regulatory framework on US-based companies could slow them down and create a competitive disadvantage against international players who can move faster with fewer constraints.

OpenAI’s Approach to Sensitive Conversations

The panel analyzes an OpenAI blog post detailing its efforts to manage sensitive conversations in ChatGPT, particularly those related to mental health. While applauding the company for addressing the issue publicly, the panelists critique the approach for its vagueness and potential shortcomings.

Key concerns raised:

Vague and Opaque Metrics: OpenAI claims a “65% reduction in harmful responses” but provides no baseline, making the number meaningless. The panelists suspect the post was heavily vetted by lawyers to be open without revealing too much substantive information.
Conflicting Corporate Messaging: This safety-focused initiative seems to clash with other statements from company leadership about “letting adults be adults” and allowing more permissive content on the platform.
Product vs. API Endpoints: There is concern that these safety measures may only be implemented at the ChatGPT product level, leaving the raw API endpoints without the same protections. This could expose users of third-party applications built on the API to greater risk.
A “Tacked-On” Solution: The strategy of routing a user to a “safe model” only after detecting distress was described as putting them in a “padded room” after a problem arises. The panel argues that if AI is to be used for companionship, safety must be a core, foundational element of the design, not an afterthought.
Questionable Interventions: The panel questions the effectiveness of the model’s prescribed intervention, which is to advise emotionally reliant users to seek more “real-world conversations.” Users may be turning to AI precisely because human interaction is difficult or unavailable for them.

Data Centers in Space: Cool Idea, Impractical Reality

Finally, the panel discusses NVIDIA’s support for a company called StarCloud, which aims to build data centers in outer space. The CEO’s bold claim is that “in 10 years, nearly all the new data centers will be built in outer space.”

The panel’s reaction is a “rollercoaster” from intrigue to dismissal:

The Compelling Advantages: The core premise is sound. Space offers virtually unlimited solar energy and an unparalleled cooling environment (the vacuum of space), addressing two of the biggest operational costs and environmental concerns of terrestrial data centers.
The Maintenance and Upgrade Nightmare: The biggest flaw in the plan is the utter impracticality of maintenance. The pace of hardware innovation is relentless, with new GPUs becoming obsolete within months. The inability to physically swap a failing card or upgrade a rack in orbit makes the concept a non-starter for cutting-edge compute.
Space Debris and Geopolitics: The panelists also point to the growing problem of space debris and the immature geopolitical and regulatory landscape governing outer space as significant hurdles.
The Final Verdict: The idea is ultimately dismissed as “stupid” and a form of “lazy offloading.” The focus, one panelist argues, should be on making terrestrial hardware and models more efficient, not on “making space the problem as well.”