Over the past few weeks, I’ve had conversations with infrastructure leaders at hyperscalers and frontier model providers who are running tens of thousands of servers and gigawatts of capacity.
Here is what I learned:
1. Full stack is about economics, leverage, and speed. At budgets in the hundreds of billions, a few points of efficiency are worth billions. The feedback loop from customer to silicon is now weeks, so suppliers have to move just as fast.
2. Own the integration and not every layer. The Cloud era stack doesn’t fit the 100x AI era. Leaders want to co-design models and silicon plug-and-play components from multiple vendors by creating an ecosystem.
3. Demand still outruns supply. Unlike the Dot-com era, few see a risk of overbuilding. Compute is allocated in real time across internal products and customers, sourced from a mix of self-built, leased, and neocloud capacity.
4. The bottleneck keeps moving. Power, compute, memory, silicon, and land each take their turn. Longer-term, it may be capital: models change every few weeks, while infrastructure lasts for decades.
5. Utilization is the hidden opportunity. A gigawatt campus runs far below capacity on average. There are many opportunities to solve problems to fix this.
For founders, the opportunity is in problems the big companies can’t solve with brute force: power and cooling, data movement, AI that accelerates hard engineering, and security for self-improving systems. Come ready to co-design.
We are still early in the AI era. As one leader put it, double compute tomorrow, and they would use all of it.
Bottom line: We are in a Golden Era for Semis and Systems. AI leaders are going full stack, and the winners will own the integration while opening the box in some areas to an ecosystem of startups.
In the news this week, Microsoft unveiled its Maia 300 inference chip with plans for multi-gigawatt deployment, TSMC reported a 69% surge in quarterly profit, and NVIDIA introduced new AI-powered semiconductor manufacturing tools. Meanwhile, Anthropic’s potential IPO valuation came under scrutiny, and Meta launched its Muse Wonder frontier models and Muse Spark Open.
Full Weekend Edition below. 👇
Signals Shaping the Future of AI:
Infrastructure
NVIDIA unveiled new AI-powered semiconductor manufacturing tools with partners including Applied Materials, Cadence, KLA, Lam Research, and Synopsys. Its CUDA-X platform is being used to accelerate chip design, simulation, and inspection, helping manufacturers improve efficiency and production capacity. Click here
Microsoft unveiled Maia 300 and plans multi-gigawatt deployment of its in-house AI inference systems starting in 2027. The company expects Maia 300 to deliver 40% more tokens per megawatt than leading GPU-based systems. Click here
TSMC reported a 69% jump in third-quarter net profit as AI chip demand continued to surge. The manufacturer maintained plans to invest $60 billion to $64 billion in 2026 capital spending. Click here
ASML forecast stronger fourth-quarter sales as demand for advanced AI chipmaking equipment remained robust. Its outlook underscores continued investment in semiconductor manufacturing capacity. Click here
NVIDIA and Cisco announced a new optical networking standard designed to quadruple bandwidth density for AI data centers. The technology targets up to 12.8 terabits per second of capacity while improving power efficiency and reducing network complexity. Click here
The competition for scarce AI compute is pushing leading labs into new infrastructure partnerships. Google, OpenAI, and Anthropic are contending with rising capacity costs and long-term supply commitments. Click here
Enterprise
Meta unveiled its Muse Wonder frontier models and Muse Spark Open, an 816-billion-parameter open-weight model. The releases advance Meta’s capabilities in coding, reasoning, and agentic tasks, with Muse Spark Open optimized for Apple’s M5 Ultra Macs. Click here
Bank of America hired 1,000 contractors to help repair software vulnerabilities surfaced by advanced AI models. The response illustrates the operational cost of turning increasingly capable AI tools on enterprise codebases. Click here
Google launched Universal Cart in the U.S., bringing shopping activity across Search and Gemini into a unified experience. Built around its Universal Commerce Protocol, the tool connects product discovery, price tracking, and checkout across retailers. Click here
Figma introduced Specs to help teams define design rules that AI agents can follow. The Markdown-based feature brings design guidance and contextual references into agent workflows. Click here
Anthropic replaced Git with an in-house source-control system after AI-driven code changes surged more than 10-fold. The company built Scion in five months as daily merges climbed from 2,500 to over 30,000, exposing new bottlenecks in software development infrastructure. Click here
Capital Flows
Anthropic reported $16.8 billion in third-quarter revenue as investors reconsidered its potential $2 trillion IPO valuation. Reuters reported that some investors discussed a valuation closer to $1.5 trillion as quarterly revenue growth slowed to 45%. Click here
Ramp raised $1.85 billion at a $60 billion valuation as it expanded its AI-enabled financial operations platform. The company reported more than $2 billion in annualized revenue and 85,000 customers. Click here
Modal raised $750 million in a Series D at a reported valuation of $15.75 billion. The AI infrastructure company said it had surpassed $1 billion in annualized revenue run rate. Click here
Blackstone-backed AirTrunk postponed its planned $1.5 billion Singapore IPO amid weakening investor appetite for AI infrastructure listings. The delay followed the withdrawal of NVIDIA-backed Firmus’s nearly $5 billion Australian IPO amid rising borrowing costs and market volatility. Click here
Research
Google Quantum AI demonstrated a 40-fold reduction in quantum computing errors using its Willow processor. The breakthrough achieved 99.99% logical fidelity using a technique called magic state cultivation, addressing a major bottleneck in scaling practical quantum computers. Click here
Anthropic described new techniques for detecting and disrupting illicit model distillation. The company said its defenses substantially reduced observed attempts to extract proprietary model capabilities. Click here
AI researcher David Silver outlined a reinforcement-learning-led approach to building systems that learn from experience. His startup Ineffable Intelligence is developing a path beyond reliance on human-generated training data. Click here
Anthropic published research showing that some smaller AI models can operate across inexpensive connected devices. The study examined the risk of decentralized self-replication without centralized data-center infrastructure. Click here
Policy
The FTC issued subpoenas to OpenAI and Anthropic as part of an investigation into AI safety and potential consumer-protection violations. The agency is seeking documents on security incidents, safeguards, and product claims. Click here
At least 8 U.S. states filed lawsuits against OpenAI over alleged consumer-protection violations involving ChatGPT. The complaints allege the company misled users about chatbot safety, including risks to children. Click here
AI data-center development on federal land emerged as a major policy issue in Nevada’s gubernatorial race. Proposed projects face legal challenges and growing scrutiny over environmental reviews, electricity costs, and local oversight. Click here
AI safety advocacy groups have contributed approximately $48 million to influence U.S. congressional races in 2026. The spending reflects a growing political battle over federal AI regulation and oversight. Click here
Global AI Strategy
China proposed new export controls on 23 AI technologies, including advanced models, AI agents, chip designs, and computing clusters. The restrictions would tighten Beijing’s oversight of strategically important AI technology transfers abroad, with some interim registration requirements taking effect immediately. Click here
The UK confirmed that its AI Security Institute retained access to pre-release models from leading U.S. AI developers. The arrangement preserves Britain’s role in independent frontier-model testing. Click here
NVIDIA committed $250 million in accelerated-computing resources for UK AI safety research over five years. The support will focus on security and agentic AI systems. Click here
SK Hynix’s Solidigm sought U.S. approval for chipmaking equipment in China as it considered a new American factory. The reported negotiations highlight the intersection of AI storage demand, export controls, and semiconductor industrial policy. 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.
Retym is building high-speed connectivity silicon for AI infrastructure, using advanced signal-processing technology to improve bandwidth, latency, and efficiency across next-generation data-center networks. The company is hiring for roles across silicon, software, validation, DSP, and systems engineering. Click here
Modal builds cloud infrastructure for AI inference, training, and agents, with a focus on reducing deployment friction and improving resource utilization. It is hiring across inference engineering, infrastructure security, systems, and forward-deployed engineering. Click here
Metaview is building AI coworkers for recruiting, expanding from interview intelligence into agents that search, screen, and handle repetitive hiring workflows so teams can focus on judgment and relationships. The company is hiring across engineering, product, sales, customer engagement, marketing, recruiting, and operations. 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.
Satya Nadella (Click here) — “We need to separate the supply of intelligence from the authority over it. The most trustworthy Super Intelligence system will not be the one with the model we trust most. It will be the one that enables us to trust the model the least.” In a post viewed 13.7M times, Nadella argues that enterprises should treat frontier models like powerful insider risks, not because they are inherently malicious, but because capable systems with access to sensitive data and mission-critical actions need strong external controls. His framework centers on observability, independent auditability, model diversity, containment, and keeping permissions and safeguards outside the model itself.
Jyoti Bansal (Click here) — “The unit of autonomy has to become larger than a coding task. An agent that writes a change and hands it off has automated one step. A software factory has to carry that change through integration, testing and release, and use failures to improve the next attempt.” Bansal argues that as more coding agents work in parallel, the bottleneck shifts from code generation to the systems that verify, integrate, and safely ship those changes. His takeaway is that the real measure of an AI software factory is not how much code agents produce, but how much correct work reaches production with minimal human intervention.
Greg Isenberg (Click here) — “The model is the CPU, the harness is the operating system, and skills and plugins are the apps. Thinking about it that way makes it easier to see where the whitespace is. The best teams I talked to treat the model as swappable.” In a post viewed 2.2M times, Isenberg shares takeaways from conversations with AI founders and executives across San Francisco, arguing that value is shifting toward data, vertical harnesses, evals, agent infrastructure, and distribution. His broader point is that as models commoditize and software becomes easier to rebuild, the durable advantage will come from owning proprietary workflows, context, and the systems around the model.
