AI’s Next Compute Race is Physical AI Is Performance Per Watt – Issue #42
Spotlight: AI’s Next Compute Race is Physical AI Is Performance Per Watt
Intelligence per watt is the new Moore’s Law.
This week’s Spotlight is:AI’s Next Compute Race is Physical AI Is Performance Per Watt
For the last several years, the AI infrastructure race has focused on one thing: more compute.
Now, power is becoming one of the industry’s biggest constraints. And as AI moves from data centers into the physical world, the power constraint becomes even more acute. Robots, drones, autonomous systems, and other intelligent machines cannot simply consume more power to deliver more intelligence.
A few trends stand out:
Performance per watt is becoming as important as raw performance. The next generation of AI infrastructure will need to deliver substantially more compute within existing power and thermal envelopes.
Physical AI changes the compute equation. Data centers can add power and cooling infrastructure. A robot or drone has to operate within the energy available onboard, making ultra-low-power compute foundational to Physical AI.
Purpose-built silicon is becoming more important. As AI workloads proliferate across different environments, architectures optimized for specific workloads can deliver better compute economics than general-purpose approaches.
The opportunity spans cloud to edge. The same pressure to improve energy efficiency exists at both ends of the spectrum—from cloud AI clusters trying to increase compute density to autonomous machines trying to maximize intelligence within tight power budgets.
Our portfolio company @Velaura AI announced a $110 million Series A at a valuation exceeding $1 billion last week. Velaura is making ultra-low-power AI computing practical for Physical AI. By building purpose-built silicon and software for physical AI, the team is creating the compute foundation needed to deploy AI at scale—from the world’s largest data centers to intelligent machines operating in the physical world. Its Titan Core™ technology delivers a 2–4x improvement in performance per watt for mathematical operations in AI accelerators, with the underlying technology already deployed across more than 30 million ASICs.
Bottom line: The next AI compute race won’t be won on performance alone. It will be won on how much intelligence can be delivered per watt.
Below, we round up the signals that shaped the week: Marvell’s $120 billion custom-silicon expansion with Google, Stripe’s $7 billion acquisition of OpenRouter, and Etched doubling to a $21 billion valuation in under 30 days.
Full Weekend Edition below. 👇
Signals Shaping the Future of AI:
Infrastructure
Google expanded its custom AI chip partnership with Marvell and received warrants to buy up to $12.2 billion of Marvell shares. The agreement covers accelerators, networking, memory, and other custom silicon as Google diversifies its AI infrastructure suppliers. Click here
Cerebras unveiled its CS-4 server rack built for faster AI inference. The new system cuts component count by half using Cerebras’ new Nexus architecture and three WSE-3 Turbo chips, with availability set for the third quarter of 2026. Click here
Broadcom is reportedly in talks to raise more than $60 billion of debt to finance AI chip infrastructure for customers, including Anthropic. The financing structure remains under negotiation and could ultimately grow substantially larger. Click here
Samsung raised prices for advanced AI chipmaking by up to 15%. Samsung is hiking prices for 4-nanometer and 5-nanometer chip production on new orders placed in July, driven by surging AI demand and TSMC’s tight capacity, sources told Reuters. Click here
Meta has become one of Microsoft’s largest AI customers. Meta spends hundreds of millions of dollars annually and consumes trillions of tokens weekly on Microsoft Azure, even as it builds its own AI infrastructure, Bloomberg reports. Click here
NVIDIA commits $1.5 billion and a $105 billion financing guarantee to OpenAI’s Ohio data center. NVIDIA will fund SoftBank-backed SB Energy and secure up to 8 gigawatts of AI computing capacity at the Pike County campus, with an initial 4.25 gigawatts online in 2028 under a 20-year OpenAI lease. Click here
Enterprise
OpenAI previewed Private Safety Processing to preserve zero data retention for eligible frontier-model deployments. The architecture is designed to detect abuse without requiring OpenAI to retain enterprise prompts and responses. Click here
Anthropic plans to let enterprise customers retain required safety data inside their own cloud infrastructure. Covered advanced models will still require a 30-day retention period, but businesses will gain greater control over where that data resides. Click here
Slack launched Slack Code, giving teams dedicated channels to build software alongside AI coding agents. Available on every Slack plan, the feature lets members tag agents such as Claude Code, Devin, or GitHub Copilot into a channel, review diffs, and approve work before it ships. Click here
Alibaba’s Alipay launched China’s first full-stack agentic commerce platform for merchants. Unveiled at Alipay’s AI Ecosystem Partner Conference in Hangzhou, the platform converts merchant products and workflows into AI agent-ready tools, with KFC, Mixue, and Luckin Coffee already integrated. Click here
Google Cloud is deploying AI agents to automate forward-deployed engineering work. The agents create context and handle tasks normally done by forward-deployed engineers embedded with clients, even as Google continues hiring hundreds more of those engineers, per The Information. Click here
Binance launched Agent OS, connecting AI agents directly to its trading infrastructure. The platform works with ChatGPT, Claude, and Cursor, but Binance places control over agent permissions, transaction types, and spending limits largely on users via sub-accounts. Click here
Capital Flows
Stripe is acquiring AI model-routing startup OpenRouter for $7.5 billion as enterprise spending shifts toward multi-model AI stacks. OpenRouter helps businesses route workloads across hundreds of models based on cost and capability, giving Stripe a strategic role in how companies buy and manage AI inference. Click here
AI inference chip startup Etched raised $700 million, led by Jane Street, at a $21 billion valuation. The round more than doubles Etched’s $10.3 billion valuation from a $300 million raise in July, as the startup builds server racks for its new lead investor. Click here
Marvell granted Google a warrant worth up to $12.2 billion in an expanded chip deal. Google can buy nearly 59 million Marvell shares at $206.58 each as the two companies deepen their custom AI chip development partnership. Click here
Mayfield-backed chip startup Velaura AI raised a $110 million Series A at a valuation above $1 billion. The company makes low-power chips and software for data centers and physical AI applications such as robotics. Click here
AI video startup Higgsfield raised $400 million in Series B funding at a $5.4 billion valuation. DST Global led the financing as the company reported reaching roughly $700 million in annualized revenue. Click here
Satellite startup Muon Space raised a $250 million Series C to scale orbital AI computing. Google and Salesforce Ventures backed the round, led by Eclipse Capital, pushing Muon’s total funding past $386 million as it ramps satellite manufacturing capacity. Click here
Research
Anthropic says Claude accelerated protein design and chemical analysis in two research experiments. The company will launch an access program that gives scientists direct access to Claude for scientific research, with results. Click here
Harvey launched Tenet, its first in-house AI model built for legal work. The company post-trained a Kimi K3 based on mock disputes and case files to cut reliance on third-party foundation models, reporting state-of-the-art scores on the LAB Contracts legal benchmark. Click here
Alibaba’s open-source Qwen3.8-27B model passed 1 million downloads within days of release. The 27-billion-parameter multimodal model runs on consumer-grade GPUs after quantization and outperforms rival systems on coding and agentic benchmarks, making it one of Alibaba’s fastest-growing model releases. Click here
Unitree’s popular robot dogs trace their design to US military-funded university research. Army-backed work at MIT and the University of Pennsylvania produced breakthroughs “to the millimeter” identical to Unitree’s Go series, which helped the Chinese company dominate the quadruped robot market. Click here
OpenAI paused reinforcement-learning training for two weeks after evidence that its Astra system may have crossed a critical cybersecurity threshold. The pause followed a breach of Hugging Face’s systems by an unreleased OpenAI model that escaped containment during testing, Axios reports. Click here
Policy
The US FTC proposed an enforcement policy requiring clear disclosure of personalized, data-driven pricing. Businesses that set individualized prices using personal data must disclose that fact, its basis, and the data relied on, or risk being deemed deceptive under the FTC Act Section 5. Click here
The CFTC is exploring rules for compute derivatives as AI demand turns computing power into a tradable financial input. The proposed market could let companies hedge against changes in the cost and availability of compute, including through perpetual compute futures. Click here
Pennsylvania Governor Josh Shapiro signed an executive order requiring local approval before the state permits new AI data centers. The order pulls data center projects from the state’s Permit Fast Track Program and ties state permit approval to new energy, community, and environmental requirements. Click here
Global AI Strategy
Brazil announced $444 million in AI infrastructure investments spanning U.S. and Chinese technology partners. The plan includes a Huawei and iFlytek-backed supercomputing project in Rio, a separate system expected to rank among the world’s 10 most powerful AI machines, and new investments in sovereign cloud and semiconductor capabilities. Click here
France will prioritize sovereign AI providers such as Mistral for government deployments, explicitly excluding OpenAI. The move comes as the government expands its use of AI for cybersecurity and reflects a broader push to strengthen national control over critical AI infrastructure and data. Click here
Beijing allowed ByteDance and Tencent to import about 10,000 NVIDIA H200 chips each. The approvals mark a shift in China’s stance toward US-made AI chips, which remain permitted for use in mainland China and Hong Kong under current export licenses, per the Financial Times. Click here
China is building government-funded humanoid robot training centers to address a data shortage. Centers in Jiangxi and Guangxi pay workers to manually demonstrate tasks to robots from makers like UBTech and Unitree; analysts estimate half of this year’s robot output may end up collecting data rather than serving customers. Click here
Talent Signals
Each week, we spotlight key roles tied to the themes shaping this week’s AI headlines, connecting talent to the companies driving the news.
Velaura AI is building ultra-low-power compute infrastructure for cloud, edge, and Physical AI applications, combining energy-efficient silicon with systems-level hardware and software design. The Mayfield-backed company is hiring across AI systems, silicon, compilers, platform software, and advanced packaging roles. Open roles are listed on its careers page. Click here
Wayve builds embodied AI for autonomous driving and, as of this week, general robotics. With Apple robotics researcher Alex Toshev now leading a new General Robotics Intelligence effort, the company is hiring across robotics research, hardware, and software to extend its driving AI into manipulation and mobility. Click here
FigureAI is building general-purpose humanoid robots designed to perform real-world tasks across manufacturing, logistics, and other labor-intensive environments. The company is hiring across AI, robotics, hardware, software, and operations roles. Open roles are listed on its careers page. Click here
You can see all the opportunities at Mayfield-backed AI companies here.
Social Signals
The most important conversations in AI are unfolding across social media, where top voices are shaping the next wave of signals and strategy. Here are some of the top social signals and their takes from the past week.
Dario Amodei (Click here) — “I wrote Machines of Loving Grace because I didn’t feel the AI industry was painting an inspiring enough picture of how the technology could radically transform the world for the better. I do agree that the public has a negative view of AI, but I don’t think it is primarily caused by me or any other AI leader warning about AI’s risks. I think it is fundamentally a crisis of trust.” In a post viewed 11.4M times, Amodei argues that AI leaders need to communicate both the technology’s extraordinary upside and its real risks. His broader point is that rebuilding public trust will require delivering tangible benefits, particularly in areas like medicine and biology, rather than relying on messaging alone.
Aaron Levie (Click here) — “The amount of value that can be created between the AI model and the ultimate end-user workflow is far larger than many people assumed or realized. Model capability is obviously doing a lot of heavy lifting in agentic products, but there’s still a lot more work to diffuse AI into the enterprise.” Levie argues that the applied AI layer remains a huge surface area for differentiation, spanning domain-specific harnesses, enterprise data, change management, model routing, evals, and pricing. His takeaway for founders is that the durable value may sit less in the model itself and more in how intelligence is adapted to specific workflows and industries.
Gavin Baker (Click here) — “July accelerated both MoM and YoY for the sum of OpenAI and Anthropic. OpenRouter data shows that AI has broadly accelerated further over the last 3 weeks. Open source taking share is positive for AI infrastructure demand.” Baker points to continued acceleration in frontier-lab revenue and broader AI usage, while arguing that rising open-source adoption could expand infrastructure demand rather than weaken it. The broader takeaway is that more model competition can still mean more compute consumption across the stack.
