🦾 From GPUs to LPUs – Where Groq Fits Among Nvidia, AMD, and Cerebras

Happy Monday! Here’s what’s inside this week’s newsletter:

• Deep dive: In light of Groq’s recent $750M funding round, we explore its Language Processing Unit (LPU), purpose-built for inference, and how the company compares to Nvidia, AMD, and Cerebras.

• Spotlights: The European Semiconductor Industry Association’s (ESIA) position paper on the proposed EU Chips Act 2 and Microsoft’s blueprint for what it calls the “world’s most powerful data center”.

• Headlines: Nvidia’s $5B Intel stake, Quantinuum’s ‘unconditional’ quantum supremacy claim, new photonic and neuromorphic advances, data center and cloud infrastructure moves, and soaring AI valuations.

• Readings: Advanced foundry revenues, memory scaling, quantum investment strategies, 3D-printed optics, neuromorphic markets, gigawatt-scale data centers, and edge AI trends.

• Funding news: A slower week overall, with activity spanning early-stage financings in quantum, photonics, and data centers, alongside larger raises in AI and networking, capped by Groq’s $750M round.

• Bonus: The latest front in the U.S.–China chip wars, as Washington and Beijing escalate export restrictions and probes, while Huawei accelerates its own AI chip and infrastructure push.

Last week, Groq raised $750M at a $6.9B valuation, more than doubling its valuation from just a year ago. Combined with a $1.5B Saudi Arabia infrastructure deal, Groq is positioned as a serious contender in the AI hardware race.

But what exactly is Groq’s bet? The answer lies in the Language Processing Unit (LPU), Groq’s custom-built chip designed specifically for AI inference.

What is a Language Processing Unit (LPU)?

The LPU is Groq’s clean-sheet alternative to GPUs, built solely for AI inference. While GPUs evolved from graphics and excel at parallel training, inference (especially real-time, batch-of-one workloads) requires a different architecture.

The LPU is built around four first-principle design choices:

• Software-first: The compiler was built before the chip. It schedules every operation deterministically across LPUs, removing the need for custom kernels and ensuring full utilization.

• Programmable assembly line: A single pipeline (“conveyor belt”) streams data through compute units in lockstep. Multiple LPUs link seamlessly, scaling linearly without external switches or routers.

• Deterministic execution: No caches, no branch predictors, and no variability. Each instruction takes a fixed number of cycles, guaranteeing predictable latency for real-time systems.

• On-chip memory: 80 TB/s SRAM bandwidth keeps data local, ~10× faster than GPUs using off-chip HBM (~8 TB/s), cutting latency and boosting efficiency.

Result: Up to 10× lower latency and 10× higher memory bandwidth than GPUs for inference.

Can Groq Overcome Nvidia’s Moat?

Nvidia is not standing still, and CUDA remains a massive moat. Beyond software inertia, enterprises calculate total cost of ownership (TCO) around Nvidia’s ecosystem, which makes switching even harder unless gains are overwhelming. But Groq has a strong tailwind: inference demand is growing faster than Nvidia can supply, and governments and enterprises seek cost savings, sovereignty, and diversification. Still, Nvidia is likely to keep leading the AI hardware market, though Groq’s focus on inference is carving out real momentum.

How Does Groq Compare to Nvidia, AMD, and Cerebras?

Groq

• Valuation: $6.9B (private, 2025)

• Performance: LPU chip (~725 TOPS); ~10× lower inference latency vs GPUs

• Efficiency: On-chip SRAM (~10× GPU bandwidth); linear scale-out without bottlenecks

• Target Market: Real-time inference (LLMs, NLP, vision); GroqCloud API & GroqRack clusters

→ Groq is purpose-built for inference, with deterministic design and on-chip memory that deliver latency and efficiency GPUs can’t match. But to succeed, it must also grow adoption of its own compiler and software stack against Nvidia’s CUDA ecosystem, which most developers rely on and rarely switch away from.

Nvidia

• Valuation: ~$4T (public, 2025)

• Performance: H100 GPU - one example of many - (~1,000 TFLOPs; 80–96 GB HBM); gold standard for training, strong inference throughput at high batch sizes

• Efficiency: Off-chip HBM; ~700 W per GPU; optimized by CUDA ecosystem

• Target Market: End-to-end AI across cloud, data centers, and edge

→ Nvidia remains the $4T giant with ~80% market share, dominating both training and inference. GPUs excel at high-throughput inference, though batch-of-one, real-time workloads remain less efficient — the exact gap Groq is trying to exploit. CUDA lock-in, however, makes it hard for customers to switch.

AMD

• Valuation: ~$150B (public, 2025)

• Performance: MI300X GPU (192 GB HBM3; 5.3 TB/s bandwidth); outperforms Nvidia H100 on LLM inference throughput

• Efficiency: Fewer GPUs per model thanks to larger memory; chiplet design improves perf/W

• Target Market: Large-model inference and HPC; enterprise AI

→ AMD’s MI300X closes the gap, with 192 GB of HBM3 and higher LLM inference throughput than Nvidia’s H100, allowing larger models to fit on fewer chips. This memory advantage reduces complexity and boosts efficiency, but adoption is still slowed by CUDA inertia despite AMD’s stronger raw performance.

Cerebras

• Valuation: ~$3B (private, 2024)

• Performance: WSE-3 wafer-scale chip (~125 PFLOPs BF16; 40+ GB on-chip memory); 850k cores on a single wafer

• Efficiency: On-wafer memory; no networking overhead; ~20 kW power per wafer

• Target Market: Ultra-large model training/inference; sovereign AI and national labs

→ Cerebras takes the opposite approach to Groq’s modular pipeline, concentrating compute on one massive wafer for extreme-scale models. Its WSE-3 delivers 125 PFLOPs and 40+ GB of on-chip memory while avoiding multi-chip networking. The tradeoff is cost and power, with each wafer drawing ~20 kW.

Sources: Groq’s $6.9B AI Chip Surge: Inside the Inference Revolution (TS2, 2025);What is a Language Processing Unit? (Groq, 2025)

🦾 ESIA Position Paper: EU Chips Act 2 — Europe’s Path Towards Semiconductor Leadership (European Semiconductor Industry Association)

The European Semiconductor Industry Association (ESIA) has published a position paper on the proposed EU Chips Act 2. Building on the 2023 Chips Act that aimed to double Europe’s market share to 20% by 2030, ESIA calls for a stronger focus on industrial deployment, AI leadership, and innovation infrastructure to close Europe’s competitiveness gap.

Key recommendations include:

• Establishing a dedicated semiconductor budget with faster and more flexible funding.

• Expanding “first-of-a-kind” (FOAK) support to cover upstream suppliers, equipment makers, and joint ventures.

• Prioritizing chips for AI, both foundational and leading-edge.

• Creating an institutionalized high-level dialogue between policymakers and industry.

• Simplifying administrative and regulatory rules to boost innovation.

💥 Microsoft unveils the “world’s most powerful data center” (Heise)

“Microsoft unveils the blueprint for its next generation of AI data centers. The first location will be built in Mount Pleasant, Wisconsin, where Microsoft plans to put hundreds of thousands of Nvidia Blackwell accelerators into operation in early 2026. The cost is expected to be $3.3 billion. According to Microsoft, the Fairwater data center in Wisconsin will provide 10 times more computing power than the world's best-equipped data center today when it is completed. It is not clear which hyperscaler will operate it – private companies do not register their systems in the Top500 list of the world's fastest supercomputers. Since Microsoft, Meta, Amazon, and Google, among others, already operate data centers with more than 100,000 accelerators, Microsoft's new building could approach the million GPU mark.”

Last week’s headlines featured new semiconductor collaborations, a dense wave of quantum breakthroughs, advances in photonic and neuromorphic tech, developments in data centers and cloud, and soaring AI valuations.

🦾 Semiconductors

Nvidia buys $5B stake in Intel, planning AI chip collaboration (TechCrunch)

Siemens unveils groundbreaking Tessent AnalogTest software for automated analog circuit test generation (Siemens)

MediaTek develops chip utilizing TSMC’s 2nm process, achieving milestones in performance and power efficiency (MediaTek)

NVIDIA and United Kingdom Build Nation’s AI Infrastructure and Ecosystem to Fuel Innovation, Economic Growth and Jobs (NVIDIA)

New AI-native processor for edge applications offers 100x power and performance improvements over 32-bit MCUs (Ambient Scientific)

Nvidia just spent over $900 million to hire Enfabrica CEO, license AI startup’s technology (CNBC)

Trump-UAE Deals: AI Chips and Investments Spark Controversy (Markets.com)

⚛️ Quantum

Quantum Motion Delivers the Industry’s First Full-Stack Silicon CMOS Quantum Computer (Quantum Motion)

Julich Supercomputing Center becomes world’s first HPC center to deploy NVIDIA DGX-Quantum System (PR Newswire)

Quantinuum Researchers Report ‘Unconditional’ Quantum Supremacy (The Quantum Insider)

IonQ Announces Intent to Acquire Vector Atomic, Expanding Into Quantum Sensing and Strengthening Its Quantum Technology Portfolio (IonQ)

Riverlane Launches DeltaKit, A New Software Platform for QEC (Riverlane)

Infleqtion Unveils New Architecture to Accelerate Its Quantum Computing Roadmap To Achieve 1000 Logical Qubits by 2030 (Infleqtion)

“Quantum squeezing” a nanoscale particle for the first time (Phys.org)

Physicists create new electrically controlled silicon-based quantum device (Phys.org)

⚡️ Photonic / Optical

LightSolver Announces Breakthrough in Physical Modeling on the LPU and New Roadmap for Optical Analog PDE Solving (LightSolver)

Nanoscale optical device enables independent control of light intensity and phase using electricity (Phys.org)

OPROCESSOR INC Announces Breakthrough VCSEL-Based Photonic Interposer Technology (OPROCESSOR)

🧠 Neuromorphic

BrainChip expands global reach; announces Akida boards and AI development kits available at DigiKey (FOX4KC)

UCL to lead UK’s brain-inspired computing push with new innovation centre (UCL)

💥 Data Centers

OpenAI announces Stargate UK, with up to 8,000 GPUs at Nscale data centers (Data Center Dynamics)

Cadence Adds Digital Twin for Nvidia’s AI Data Center Compute Platform (All About Circuits)

Axiom Space, Spacebilt Announce Orbital Data Center Node Aboard International Space Station (Axiom Space)

If you’d like to learn more about orbital data centers, check out our interview with Starcloud!

☁️ Cloud

How AI Startups Are Fueling Google’s Booming Cloud Business (TechCrunch)

🤖 AI

UK at risk of becoming AI “users, not makers”, says Hoberman, as Nvidia and Microsoft pledge billions (Tech.eu)

Elon Musk’s xAI raising $10 billion at $200 billion valuation: sources (CNBC)

This week’s reading list includes updates on advanced foundry revenues and memory scaling, quantum investing and machine learning, 3D-printed optics, neuromorphic theory and markets, the rise of gigawatt-scale data centers, and edge AI and connectivity trends.

🦾 Semiconductors

Global Semiconductor Foundry 2.0 Market’s Q2 2025 Revenue Up 19% YoY Driven by Advanced Process and Packaging (Counterpoint Research) (5 mins)

Scaling Memory With Molybdenum (SemiEngineering) (7 mins)

⚛️ Quantum

How to invest in quantum stocks: A guide to long-term investing in quantum technology (The Quantum Insider) (15 mins)

Quantum Optical Circuits Enable Kernel Learning for Support Vector Machines (Quantum Zeitgeist) (10 mins)

⚡️ Photonic / Optical

3D-Printed Optics and Photonics: Unlocking the Third Dimension (Optics & Photonics News) (10 mins)

Photonics Market by Product Type ‒ Global Forecast to 2030 (Markets and Markets) (12 mins)

Photonics Chipmakers Race to Production (EE Times) (13 mins)

Semiconductors & Photonics (2025) (Sifted) (20 mins ‒ Paywall)

🧠 Neuromorphic

Neuromorphic Intelligence Leverages Dynamical Systems Theory to Model Inference and Learning in Sustainable, Adaptable Systems (Quantum Zeitgeist) (13 mins)

Neuromorphic Hardware Market Revenue and Forecast by 2033 (Precedence Research) (5 mins)

How does negative differential resistance relate to neuromorphic computing and sensors? (EEWorldOnline) (4 mins)

💥 Data Centers

xAI’s Colossus 2 – First Gigawatt Datacenter In The World (SemiAnalysis) (10 mins)

Data Center Server Rack Market Size, Share and Trends 2025 to 2034 (Precedence Research) (60 mins)

Five charts reveal how AI drives soaring data-center energy use and emissions (RECCESSARY) (18 mins)

📡 Connectivity

Private 5G networks and the Edge opportunity for telcos (Data Center Dynamics) (24 mins)

🤖 AI

From ARM to Edge AI Disruptor: How Noel Hurley Is Leading the Change with Logic-Based AI at Literal Labs (Tech.eu) (7 mins)

Making the Case for a Third AI Technology Stack (Brookings) (15 mins)

Last week saw fewer rounds than the week before, a little over half in number. Activity ranged from early-stage financings in quantum, data centers, and photonics to larger raises in AI and networking, capped by Groq’s $750M round.

And again: The U.S. and China are tightening the screws on each other’s semiconductor industries. Washington added two Chinese chipmakers to the so-called Entity List, while Beijing barred companies from buying Nvidia’s AI chips, accused it of antitrust violations, and launched new probes into U.S. semiconductors. Meanwhile, Huawei is moving fast to fill the gap with its own AI infrastructure and chip plans.

Huawei Announces New AI Infrastructure as Nvidia Gets Locked Out of China (TechCrunch)

China tells its tech companies they can’t buy AI chips from Nvidia (TechCrunch)

China Finds Nvidia Violated Anti-Monopoly Law, Will Carry Out Further Investigation (CNN)

China launches probes targeting US semiconductors ahead of Madrid trade talks (AP)

US penalizes two Chinese companies that acquired tools for chipmaker SMIC (Reuters)

China’s Huawei hypes up chip and computing power plans in fresh challenge to Nvidia (Reuters)

Nvidia CEO says he’s ‘disappointed’ after report China has banned its AI chips (CNBC)

❤️ Love these insights? Forward this newsletter to a friend or two. They can subscribe here.