Google’s quiet AI win spells trouble for Amazon | Official Website of Louis Velazquez, entrepreneur, finance guy and tech innovator

At first, it felt like Google (GOOGL) was caught flat-footed in the AI arms race.

Bard’s rocky debut and lackluster chatbot responses had everyone practically writing off the search giant in the AI race.

💵💰Don’t miss the move: Subscribe to TheStreet’s free daily newsletter💰💵

Fast forward a few years, and Google’s been busy behind the scenes, quietly developing a next-gen arsenal.

Now it’s rolling out custom Tensor Processing Units (TPUs), launching slick new AI models, and landing some of the who’s who in the AI space as customers.

That early stumble might have just been a warm-up for Google stock’s massive AI comeback.

Lately, Google’s hybrid model has been turning heads, and ironically, it just got the nod from an old AI rival, completing a full-circle flex.

Google stock scores chip win as Amazon lags.

Image source&colon; Morris/Bloomberg via Getty Images

Inside Google’s TPU push: faster chips for AI’s next wave

Google has been a software-first company since the outset and has efficiently woven AI into every layer of its software stack. Its push into custom AI chips, though, highlights that it can thrive on the hardware side, too.

Think of Google’s TPUs as tailor-made silicon that can effectively handle the monster math that runs modern AI models. These application-specific integrated circuits (ASICs) are specifically designed so neural networks can hum along at full throttle.

TPU v3, for example, can train BERT, Google’s game-changing language model, roughly eight times quicker than Nvidia’s older V100 GPU.

TPU v4 took things up a notch, squeezing 1.2 to 1.7 times more performance per watt than Nvidia’s A100.

These come with a caveat, though, as Google keeps TPUs locked inside its cloud, so developers need to adapt to Google’s runtime instead of Nvidia’s more universal CUDA ecosystem.

That hasn’t stopped the big hitters like Apple, Cohere, and Anthropic from tapping TPUs for serious AI workloads.

More AI Stock News:

Sure, Nvidia still dominates the 70% to 95% share of the AI accelerator market, but TPUs are carving out a niche for high-volume inference jobs, which bank on speed and efficiency. Nevertheless, Google still makes the lion’s share of its income (roughly 80%) from ads, powering search, YouTube, and partner sites.

However, the real growth engine these days is Google Cloud, and tools like BigQuery and Vertex AI allow businesses to crunch data and build smarter stuff.

Now, Google’s leaning into hardware, too, and that massive hybrid play could pull in billions in incremental revenues, challenging Nvidia’s GPU stronghold.

OpenAI’s surprising move backs Google’s AI chips in a big way

OpenAI is sending a big signal to the AI hardware space, and it’s not good news for Amazon (AMZN) .

The Microsoft-backed AI behemoth is looking to ink a deal where it runs part of its surging workloads on Google’s TPUs. Morgan Stanley calls this move a “significant endorsement” of Google’s AI hardware.

OpenAI has predominantly tapped Nvidia’s powerful GPUs to train and run ChatGPT and its other large language models (LLMs) for years.

However, with supply bottlenecks and cost constraints, OpenAI is looking to diversify its suppliers and meet the growing demands of its customers.

OpenAI won’t be getting Google’s latest, top-shelf TPUs, but that barely matters. Perhaps the bigger flex is that OpenAI picked Google’s older chips over Amazon’s Trainium, a major blow to the online retail giant.

If this deal holds, OpenAI would be running its workloads on Google Cloud, Azure, Oracle, and CoreWeave, leaving out Amazon.

For Google, that’s a massive win for its custom silicon dreams. Also, with Apple and Cohere already on board, OpenAI’s nod will likely boost faith in Google’s AI chips.

On the flip side, Amazon’s homegrown AI chips sound good on paper, but they haven’t quite delivered.

Inferentia (inference chips) promised major cost savings, but barely powers 3% of Amazon Web Services (AWS) AI work.

Trainium (training workloads) undercuts Nvidia’s prices but keeps tripping over networking bugs and clunky software.

Hence, AWS is far from cracking Nvidia’s iron grip, despite shelling out over $4 billion on Anthropic and having Trainium 2 on deck.