Google's Gemini compute is set to crush GPT-4 by 5x

While everyone was busy watching OpenAI and Microsoft redraw the tech landscape, most people forgot about the slumbering giant sitting on the world's biggest pile of chips.

Semiconductor research firm SemiAnalysis just dropped a massive reality check on the AI industry. According to their estimates, Google has fully woken up and is pouring infrastructure into training Gemini at a breakneck pace. By the end of this year, Gemini’s pre-training FLOPS are projected to smash GPT-4 by 5x. By the end of next year, that gap could widen to an eye-popping 20x.

Anyone counting Google out of the frontier model race might want to look at the hardware math again.

Silicon Valley's brutal split: GPU-Rich vs. GPU-Poor

The report divides the AI universe straight down the middle into the GPU-rich elite and the GPU-poor underdogs.

The GPU-rich club is tiny: OpenAI, Google, Anthropic, Meta, Inflection, and X. These players command fleets of 20,000 or more A100 and H100 GPUs. In the Bay Area, ML researchers are literally using GPU allocations as bragging rights and recruiting leverage.

Then you have the GPU-poors: a long tail of scrappy startups and open-source developers struggling to scrape together enough VRAM. SemiAnalysis pulls no punches here, arguing that most of these underdogs are burning time on counterproductive dead ends.

Instead of tackling sparse MoE architectures or speculative decoding, they spend weeks trying to fine-tune dense models on underpowered gaming rigs. They obsess over public leaderboards with silly model names, optimizing for style rather than real-world accuracy, while ignoring the massive pre-training datasets required to build genuinely competitive tools.

Even respected infrastructure firms like HuggingFace, Together, and Databricks are lumped into the GPU-poor camp. Despite having world-class talent, their compute capacity remains orders of magnitude behind what is needed to challenge the frontier.

The only player that can break Nvidia’s grip

Right now, the entire AI industry is essentially acting as a capital conduit straight into Nvidia’s bank account. To make matters worse, Nvidia is eating everyone's lunch with its own DGX Cloud, poaching enterprise customers right under the noses of cloud providers.

SemiAnalysis argues there is only one force capable of saving the industry from Nvidia dependence: Google.

Google doesn’t have to play Nvidia's game because it owns its own silicon stack. Beyond its massive pool of GPUs, Google is aggressively ramping up TPUv5, codenamed Viperfish. Even without counting their TPUv4 fleet or the inference-focused TPUv5e chips, Google’s hardware buildout gives them unmatched compute density and operational efficiency for training models like Gemini.

Why it matters

This analysis shatters the romantic myth that scrappy open-source projects can outmaneuver frontier labs simply through clever community tweaks. Frontier AI is heavy capital-intensive industry. If smaller teams keep wasting precious compute on leaderboard chasing instead of meaningful engineering like speculative decoding or efficient serving, they will simply get left behind.

For enterprise buyers and developers, the takeaway is clear: foundational breakthroughs will remain concentrated in a handful of well-capitalized labs. Betting on cheap self-hosted base models to match next-generation systems is a losing strategy.

As for the frontier showdown, a 5x or 20x compute advantage would theoretically allow Google to leapfrog GPT-4 overnight. But we believe Google’s real bottleneck has never been compute—it is corporate hesitation. Google had Transformer and speculative decoding first and fumbled the rollout. If Mountain View can muster the stomach to release Gemini without neutering its capabilities to protect search ads, the AI race will look very different by year-end.