Groq secures $6.9B valuation as LPU inference outpaces NVIDIA

Financial growth and revenue

Groq raised $750 million recently, which pushed its valuation to $6.9 billion. This more than doubles its valuation from just 13 months ago when it reached $2.8 billion. Disruptive led the recent round with support from Neuberger Berman, Deutsche Telekom Capital Partners, and a large U.S.-based West Coast mutual fund manager. The round also included investments from Cisco Investments and Samsung Catalyst Fund. A $1.5 billion commitment from Saudi Arabia in February facilitates expansion. Groq expects these Saudi contracts to generate $500 million in revenue this year. The company plans to deploy 108,000 LPUs manufactured by GlobalFoundries by the end of Q1 2025. To meet developer demand, Groq intends to deploy more than 100,000 additional LPUs into GroqCloud. While the company targets 50 percent of global inference compute capacity, it faces significant competition. The startup powers AI applications for over two million developers. Mohsen Moazami, President of International at Groq, leads commercial efforts with partners including Aramco Digital and Earth Wind & Power. The leadership team includes Stuart Pann, formerly of HP and Intel, and technical advisor Yann LeCun from Meta. Groq has raised over $3 billion to date to build its global infrastructure and recently expanded with a European data center in Finland. This brings the total cash reserves to over $2 billion.

LPU latency and throughput

My verdict is that Groq’s specialized architecture beats NVIDIA on inference speed and efficiency. I find the performance gap between Groq and NVIDIA unmistakable. Because Groq LPUs process 500 to 750 tokens per second on Llama 3 70B models, they outpace the competition. NVIDIA H100 GPUs manage 10 to 30 tokens per second for the same workload. The H200 provides 15 to 40 tokens per second. Groq’s LPU achieves 20.9 ms/token for 66B models, which is 1.37x faster than GPUs with equal device counts. Groq 3 LPUs provide 315 petaflops of FP8 compute at rack scale and 640 TB/s of chip-to-chip bandwidth. The LPU uses 128 GB of on-chip SRAM to achieve 35x higher throughput per megawatt compared to Grace Blackwell NVL72 at 400 tokens per second. Groq’s compute model is tensor-first, where the fundamental unit of work is a 320-byte vector. Groq utilizes Attention-FFN Disaggregation to split the transformer along its natural fault line between attention and feed-forward network blocks. The LPU’s MEM Block provides a flat, uniformly addressable SRAM memory hierarchy. Groq’s deterministic execution ensures every operation happens in a predictable timeframe through a compiler that statically schedules every operation and data movement. The plesiosynchronous interconnect allows chips to sync like clockwork. However, the single-chip SRAM limit of 230MB means large models like Llama 70B require 500 or more chips. Llama 3.3 70B from Meta AI holds an ECI of 127.34.

Cloud pricing and the capacity crunch

Cloud GPU prices changed drastically in 2026. The increase in H100 contract prices from $1.70 per hour in October 2025 to $2.35 per hour in March 2026 demonstrates how supply shortages for older hardware affect the costs for all major cloud providers in the current global AI computing market today. April 2026 saw Blackwell spot rental prices climb to $4.08 per GPU hour, a 48 percent rise from two months earlier. By July 2026, B200 tier prices rose from $3.81 to between $4.90 and $5.64 per hour. You should watch how different provider tiers handle these shifts. Long-tail providers offer on-demand rates that are 50 to 60 percent below hyperscaler on-demand pricing. Specialty GPU clouds like CoreWeave and Lambda charge roughly 40 percent less than hyperscalers. For predictable workloads, reserved capacity cuts 30 to 50 percent off the hourly rate. At the long-tail tier, reserved 3-year H100 capacity runs between $1.30 and $1.80 per hour. The price hike reflects a shortage in packaging, memory, and power. The shortages in CoWoS technology, HBM3e memory, and datacenter cooling drive these spikes. The 2.5x pricing spread between providers determines if your infrastructure bill is reasonable. Can providers maintain enough inventory to prevent these price spikes from becoming permanent?