Happy Thursday!
Here's what to expect in this week's newsletter:
Spotlights: NVIDIA unveils the Vera Rubin platform with seven new chips and $1 trillion in orders, six industry heavyweights form an open optical interconnect standard, and Quantinuum demonstrates 94 logical qubits from 98 physical qubits
Funding News: A $200M AI networking stealth launch, a new cooling unicorn, and strategic NVIDIA investments totaling over $100B
Bonus: Why and how the Iran conflict matters for chip production
Spotlights
NVIDIA Vera CPU Rack (Credit: NVIDIA)
Jensen Huang used his GTC keynote to launch the Vera Rubin platform. It is not a single chip, but a unified system of seven: Rubin GPU, 88-core Vera CPU, NVLink 6 Switch, ConnectX-9, BlueField-4, Spectrum-6, and the Groq 3 LPU. Together, NVIDIA claims 10× the performance-per-watt of Grace Blackwell. The combined Blackwell and Vera Rubin order backlog is $1 trillion through 2027.
The Groq 3 LPU is the first product from NVIDIA's $20B acquisition of Groq. It packs 500 MB of on-chip SRAM for ultra-low-latency inference and is fabricated by Samsung. The Vera CPU enters the data center CPU market as a standalone product, directly challenging Intel and AMD. NVIDIA also previewed what comes next: a "Kyber" rack architecture (Vera Rubin Ultra, 2027) and "Feynman" as the next GPU architecture after Rubin.
Alongside the product launches, Huang signed a $100B letter of intent with OpenAI for 10 GW of Vera Rubin deployment and invested $2B in Nebius Group for a hyperscale AI cloud.
Six of the most influential AI infrastructure companies formed the Optical Compute Interconnect (OCI) Multi-Source Agreement. Its purpose is to define an open optical PHY specification for AI scale-up interconnects. The spec starts at 200 Gbps per direction and scales to 3.2 Tbps per fiber, supporting pluggable, on-board, and co-packaged optics.
The announcement landed during OFC 2026, which drew 16,000 attendees and a sold-out exhibition floor. Lightmatter demonstrated a record 1.6 Tbps per fiber in collaboration with Qualcomm. Lambda announced one of the largest CPO deployments to date: 10,000+ Blackwell Ultra GPUs with NVIDIA Quantum-X800 InfiniBand Photonics.
Last week, we covered NVIDIA's $4B investment in Coherent and Lumentum. This week, the rest of the industry followed with an open standard for optical AI interconnects.
Quantinuum demonstrated quantum computations using up to 94 error-detected logical qubits and 48 error-corrected qubits from just 98 physical qubits on a trapped-ion processor. Logical gate error rates reached approximately 1 error per 10,000 operations. That is significantly below the processor's raw hardware error rate, indicating that the error correction is actively improving reliability rather than adding overhead.
The result was achieved on Quantinuum's Helios processor, which also implemented high-rate Iceberg Codes during the same week. Separately, Quantinuum opened an R&D center in Singapore, its first hardware deployment outside the US, with plans to deploy a Helios system in-country later this year.
This matters for the IPO timeline. Quantinuum filed its confidential S-1 in January at an estimated $20B+ valuation, and results like these strengthen the case ahead of a public offering.
Headlines
Semiconductors & AI Hardware
Intel Xeon 6 selected as host CPU for NVIDIA DGX Rubin NVL8 — announced the same day NVIDIA launched its competing Vera CPU
Samsung expands NVIDIA foundry role — fabricates Groq 3 LPU; NVIDIA teases "Feynman" next-gen GPU architecture, potentially using Intel EMIB packaging
SK hynix chairman warns wafer shortage may last until 2030; company reviews potential US ADR listing
SemiAnalysis: "The Great AI Silicon Shortage" — AI takes ~60% of TSMC N3 output (projected 86% by 2027), crowding out smartphones and CPUs
Meta unveils four-generation MTIA custom chip roadmap — MTIA 300 through 500, all RISC-V, ~6-month release cadence with Broadcom and TSMC
AMD and Celestica announce "Helios" rack-scale AI platform — open-standards OCP form factor, available late 2026
Micron plans second chip facility at newly acquired Taiwan site, expanding DRAM and HBM capacity
Quantum
Xanadu negotiates CAD $390M government funding from Canada and Ontario for quantum semiconductor manufacturing ahead of SPAC merger vote
IBM publishes first blueprint for quantum-centric supercomputing — reference architecture integrating quantum alongside GPUs and CPUs
Infleqtion delivers UK's only operational 100-qubit system to NQCC; 99.73% two-qubit gate fidelity
QphoX launches first commercial quantum transducer; IBM as first integration partner
Riverlane publishes QEC roadmap targeting utility-scale quantum 3–5 years earlier than current timelines
Data Centers & Cloud
AWS and Cerebras announce disaggregated inference partnership — Trainium handles prefill, Cerebras WSE-3 handles decode; claims 5× throughput; first hyperscaler to commit to Cerebras hardware
Oracle Q3: Cloud infrastructure revenue surges 84% to $4.9B; remaining performance obligations explode to $553B; raises FY2027 guidance to $90B
US data center construction surpasses office construction for the first time; January 2026 set a record $25.2B in DC construction starts
NVIDIA and OpenAI sign $100B letter of intent for 10 GW of Vera Rubin deployment — separate from OpenAI's $110B February raise
Photonics & Networking
CEA-Leti and NcodiN partner to industrialize 300mm silicon photonics — "world's smallest laser on silicon" (500× smaller than standard)
Geopolitics & Policy
Iran conflict threatens semiconductor supply chain — $200B+ wiped from Samsung and SK hynix; risks to helium (Qatar) and bromine (Israel/Jordan) supplies
China mandates 50% domestic equipment in new fabs; record 421 orders for domestic lithography machines in 2025
DIGITALEUROPE calls for €200B semiconductor investment in Chips Act 2.0 position paper ahead of EU proposal expected ~March 25
Funding News
GTC week was dominated by strategic investments rather than traditional VC rounds. NVIDIA deployed over $100B in letters of intent and direct investments across OpenAI, Nebius, and Thinking Machines Lab. On the startup side, Eridu's $200M stealth launch and Frore's unicorn milestone stood out.
Amount | Name | Round | Category |
|---|---|---|---|
$200M+ | AI Networking | ||
$143M | DC Cooling | ||
$30M | Power Delivery | ||
$20M | GPU Optimization | ||
$12.2M | GPU Profiling |
Bonus: The Gas That Keeps Every Chip Factory Running
On March 2, Iranian drone strikes forced QatarEnergy to halt production at Ras Laffan Industrial City. That facility produces roughly 30% of the world's helium. It has been offline for 15 days, and spot helium prices have since doubled.
Today, helium is used at hundreds of points across a modern fab. The most critical: wafer backside cooling. During plasma etching, CVD, and ion implantation, wafers sit on electrostatic chucks. Helium fills the microgaps between wafer and chuck, transferring heat with roughly 6× the thermal conductivity of nitrogen. Without it, temperatures drift, yields collapse, and production stops. There is no substitute. Nitrogen is too slow. Hydrogen is flammable.
Helium also cools EUV lithography systems (ASML's machines draw over 1 MW of power), serves as a carrier gas in deposition, and is the standard for leak detection in vacuum chambers. Semiconductor manufacturing has overtaken MRI machines as the world's largest consumer of helium.
The supply concentration is striking. The US and Qatar together account for 75% of global helium production. South Korea imported 65% of its helium from Qatar in 2025. South Korea alone produces roughly 70% of the world's DRAM. All of Qatar's helium exports transit the now effectively closed Strait of Hormuz.
Advanced fabs typically hold 2–4 weeks of helium inventory. South Korean chipmakers, burned by a 2022 shortage, reportedly stockpiled enough to cover six months' worth. TSMC says it does not expect a near-term impact. But if the outage extends into April, fabs without deep reserves will face production cuts. The first processes affected would be etch and deposition, the highest-volume consumers of helium in a fab.
Unlike oil, helium has no strategic government reserve (the US sold its Federal Helium Reserve to Messer in 2024), no substitute in its core applications, and no way to bring new production online quickly. Recycling helps — leading fabs recover 60–95% — but it cannot close the gap when a third of primary supply disappears overnight.
