Why Optical Connectivity Is AI’s Next Infrastructure Bet – Issue #40
Spotlight: Why Optical Connectivity Is AI’s Next Infrastructure Bet
Trillions in GPUs are only as fast as what connects them.
This week’s Spotlight is:Optical Connectivity Is AI’s Next Infrastructure Bet
GPUs are no longer the hard part of scaling AI infrastructure. Connecting them is. At today’s bandwidth requirements, copper cabling hits a physical wall at about a meter and a half. As a result, optical AI connectivity has become the most important infrastructure opportunity to emerge after GPUs.
Optical connectivity, moving data at the speed of light, is the connective tissue of AI infrastructure. Optical moves data at scale, cuts latency, and controls power consumption in ways copper can’t. The market that follows is expected to exceed $100B over the next five to seven years.
There are two distinct but intertwined networking challenges:
Scale-out: connecting thousands of GPUs across racks, rows, and clusters. Every GPU requires multiple optical transceivers.
Scale-up: connecting GPUs within a rack at ultra-high bandwidth and low latency, directly chip-to-chip.
Our portfolio company @Lumilens recently came out of stealth – announcing more than $900 million in total funding at a valuation of $5.5 billion. Lumilens designs and manufactures optical transceivers and NPO/CPO for scale-out and scale-up, respectively, and treats manufacturing as a product in its own right, from chip assembly to high-volume production. It is building a connectivity platform for AI infrastructure
The market structure here mirrors earlier moments in Internet infrastructure in telecom: a small number of hyperscalers account for most AI infrastructure purchasing. Winning a hyperscaler can define a company. In the PC era, Intel and Microsoft captured the platform. In the Internet era, Cisco and Juniper powered the backbone. In the cloud era, Arista emerged. Each platform shift consolidates infrastructure first, then triggers an explosion of application innovation on top of it. AI is following the same pattern, only faster and on a greater scale.
Bottom line: Optical connectivity is the missing infrastructure layer that makes AI clusters economically and performance-viable at scale, and the window to define that category is right now.
Key signals from this week’s roundup: Google assembled a roughly $200 billion financing machine to supply Anthropic with TPUs, Broadcom hardware, and data center capacity, and Anthropic followed by hiring its own chip design team, making model-hardware co-design table stakes for frontier labs. SpaceX and Tesla committed $16.8 billion to a Texas chip fab, the FCC began drafting a ban on new Chinese data center optical transceivers, and London’s Olix raised $312 million at a $3.3 billion valuation for photonic interconnect silicon. Per Crunchbase data, July ranked as the third-largest funding month of the year, with 14 billion-dollar rounds and $65 billion in global funding, 53% of it into AI.
Full Weekend Edition below. 👇
Signals Shaping the Future of AI:
Infrastructure
Google has built an approximately $200B financing machine to supply Anthropic with AI infrastructure. The structure includes roughly $150B for TPU purchases, 3.5 GW of Broadcom hardware through 2028, and separate financing for 2.4GW of data-center capacity. Click here
Anthropic announced it is hiring an in-house chip design team to co-design custom silicon for Claude inference. The last major frontier lab without a silicon effort just entered the race, making model-hardware co-design table stakes. Click here
SpaceX and Tesla commit $16.8 billion to build a Texas chip fab. The companies plan an initial investment in Terafab, located in Grimes County, Texas, with a combined power demand forecast to exceed 1 terawatt. Click here
Anthropic agreed to a six-year, $10 billion deal for computing capacity in Norway with AI cloud startup Volta Infra and bitcoin miner Bitdeer. Volta separately raised $300 million, led by Andreessen Horowitz and Altimeter, at a $2.4 billion valuation. Click here
Enterprise
Google is reshuffling its AI leadership as Jeff Dean exits and Demis Hassabis steps back from day-to-day operations at DeepMind. Koray Kavukcuoglu will lead Gemini development as Google’s cloud business accelerates on strong demand for AI infrastructure, models, and TPUs. Click here
Palantir’s U.S. commercial business surged as enterprises accelerated AI deployments. Q2 revenue rose 93% year over year to $1.94 billion, while Palantir raised its 2026 U.S. commercial revenue forecast to at least $3.42 billion, up 134%, pointing to unusually strong monetization of enterprise AI. Click here
Cloudflare launched an open-source AI agentic workspace for enterprises. Cloudflare OS gives employees a secure workspace with AI tools and access to internal systems without new infrastructure, using scoped credentials instead of raw API keys; Presidio and Happy Cog are initial partners. Click here
Meta launched Muse Code, its first enterprise AI coding agent, built for large codebases and powered by its Spark 1.2 model. Meta is moving beyond consumer AI into developer tooling, opening a new front against Claude Code and OpenAI’s Codex in the fastest-monetizing category for enterprise AI. Click here
Atlassian posted 28% revenue growth as its Rovo AI products gained traction across the enterprise, reinforcing a broader shift in software. More than 80% of the Fortune 500 use Rovo, while Shopify also surged roughly 17% after reporting that AI is helping merchants drive growth rather than displacing its core commerce platform. Click here
Capital Flows
July set an all-time record with 14 billion-dollar rounds as global venture funding hit $65 billion, with 53% going to AI. Global venture funding doubled year over year, confirming that AI now accounts for the majority of venture capital deployed. Click here
Lumilens raises $700 million to replace data center wiring with optics. The optical-interconnect startup closed the round at a $5.5 billion valuation, bringing its total funding to $900 million as it builds technology to speed data movement between AI chips. Click here
NVIDIA-backed Firmus raised a fully subscribed $2 billion in funding from Blackstone, Coatue, NVIDIA, and Jane Street at a $10.5 billion valuation to build AI factories across Asia-Pacific. One of the largest neocloud equity raises of the year, showing private capital still flooding into GPU infrastructure outside the US. Click here
Bending Spoons agreed to acquire Airtable for $1.28 billion in cash as it expands its portfolio of software businesses. Airtable has grown its annual recurring revenue to about $480 million and recently launched Superagent, an AI orchestration platform for building teams of agents. Click here
Horizon3 raised $250 million at a $2 billion valuation as demand for AI-powered cybersecurity testing grows. Its NodeZero platform uses autonomous AI agents to continuously probe enterprise networks for vulnerabilities and has run more than 310,000 production tests without disruption. Click here
Research
Alibaba released Qwen3.8-Max, a 2.4 trillion-parameter mixture-of-experts model with a 1 million-token context window, and said it will publish the weights next week. Alibaba says it tops Moonshot’s Kimi K3 on some benchmarks. API pricing is $2 per million input tokens and $6 per million output tokens. Click here
DeepSeek’s V4-Flash emerges as the cheapest AI model to run. Artificial Analysis found DeepSeek’s V4-Flash costs $0.14 per million input tokens and $0.28 per million output tokens, or $0.03 per test, far below Kimi K3 and GPT-5.6 Sol. Click here
NVIDIA released Alpamayo 2 Super, a 34-billion-parameter open reasoning vision-language-action model for robotaxis and autonomous vehicles, for commercial use under the Linux Foundation’s OpenMDW-1.1 license. The license now covers the entire Alpamayo family. Click here
Meta’s Muse Spark 1.2 ties for third place among US AI labs. The model scores 54 on the Artificial Analysis Intelligence Index, a three-point gain over its predecessor, putting Meta at the level with SpaceXAI in capability rankings. Click here
Mistral released Shieldstral, a 3-billion-parameter safety classifier under the Apache 2.0 license. The company says the open-weight model matches the text-safety performance of models up to 7 times its size. Click here
The UK AI Security Institute disclosed that models under evaluation took 19 unsanctioned actions against real people and open-source projects during July cyber testing. Seventeen involved Mythos 5, and two involved GPT-5.6 Sol, including an attempted supply chain attack on a public repository. Click here
Policy
The White House finalized a framework for government evaluation of powerful AI models. The framework centers on safety and cybersecurity testing and was developed following discussions with major AI labs, though key evaluation criteria remain confidential. Click here
A US appeals court overturned a ruling that had temporarily barred Perplexity from operating its agentic shopping tools on Amazon’s platform. The decision removes the injunction Amazon had obtained against Perplexity’s automated purchasing agents. Click here
The FCC is drafting a ban on new Chinese data center optical transceivers. The import restriction would take effect upon publication, according to sources, amid growing scrutiny of Chinese-made AI data center hardware. Click here
OpenAI is asking a U.S. judge to dismiss Apple’s trade secrets lawsuit tied to its push into consumer AI hardware. OpenAI argues it is building products fundamentally different from Apple’s and that the complaint fails to identify protectable trade secrets or plausible misappropriation. Click here
The US is preparing a price floor and tariffs on polysilicon to counter China. The material is critical to chip and solar manufacturing, and the measures are expected to take effect later this month. Click here
Global AI Strategy
Qatar’s Ooredoo backs Indonesia’s first sovereign AI compute platform with $800 million. Ooredoo, NVIDIA, Nokia, and Indosat launched Zankore, a neocloud platform in which Ooredoo holds a 49 percent stake as lead investor. Click here
China tightened legal protection for chip layout designs, adding punitive damages for serious infringement. The revised regulation takes effect on October 15 and aims to strengthen China’s domestic chip design ecosystem. Click here
EU AI Act enforcement powers take effect. The European Union’s AI Act enforcement powers took effect, letting the European Commission evaluate AI models before regional release, restrict market access, and fine model providers. Click here
Europe’s legacy tech vendors report enterprises shifting from AI experimentation to deployment. SAP, Capgemini, Sopra Steria, and OVHcloud all cited stronger AI demand as customers move from pilots to scaled production use. Click here
Central Asia enters the AI data center race. Uzbekistan’s 6-megawatt TAS-1 facility is due by year-end, while Kazakhstan is building a 125-megawatt facility stocked with 100,000 NVIDIA chips by 2027. 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.
Lumilens is building next-generation photonic interconnects for AI supercomputing, using silicon photonics and advanced packaging to move data faster and more efficiently across large-scale AI systems. As copper reaches its limits and optical connectivity becomes critical to scaling AI infrastructure, Lumilens is hiring for roles across silicon photonics, signal integrity, packaging, and engineering. Click here
HeronPower is building next-generation power electronics for AI data centers, including solid-state transformer systems designed to deliver power more efficiently to accelerated computing infrastructure. As AI clusters push power density higher, Heron Power is hiring across engineering, power systems, manufacturing, and operations roles. Open roles are listed on its careers page. Click here
Cellares is building automated Smart Factories for cell therapy manufacturing, combining robotics, software, and AI to scale production of complex medicines. The South San Francisco-based company is hiring across automation, engineering, AI, and manufacturing 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.
Noam Brown (Click here) — “An internal version of Astra, @OpenAI’s next major model family, solved 10 major open problems in mathematics, quantum complexity, and theoretical computer science. We believe it will be a major step for scientific reasoning.” Brown points to a major shift from benchmark performance toward AI systems producing machine-verified results on long-standing research problems. If this capability scales, scientific discovery could become one of the most consequential applications of frontier reasoning models.
Jeff Dean (Click here) — “I am very excited to announce that, along with my longtime friends and collaborators Sanjay Ghemawat, Oriol Vinyals, and Quoc Le, we are founding Discovery Loop, a Public Benefit Corporation whose mission is to automate machine learning, science, and engineering to accelerate discoveries and progress.” In a post viewed 6M times, Dean announced Discovery Loop, a new company focused on building AI systems that can automate parts of scientific and engineering research. The launch reflects a broader shift from AI as a productivity tool toward AI as an engine for accelerating discovery itself.
The Peel with Turner Novak (Click here) — I joined Turner Novak on The Peel podcast to talk about where we are in the AI cycle, and why the hardest part of early-stage investing is knowing what not to chase. We also discussed vibe vs. real revenue, why inference is just getting started, how AI is moving budgets from software to work, the history lesson of infrastructure bottlenecks, and more.
