Everyone thinks they are a comic. And everyone in big-cap high tech says they can design better and/or cheaper AI chip alternatives to the industry leader, Nvidia. Turns out, it’s simply not so easy.
Apple and AWS have recently faced setbacks in AI growth, and Tesla has abandoned its own Dojo Supercomputer chip development, stating that it will switch to Nvidia and AMD for training AI models. (Like many semiconductor developers, Nvidia is a client of Cambrian-AI Research). Oh, and today, The Information reported that “Microsoft’s AI Chip Effort Falls Behind.” There is definitely a critical trend here.
A few companies have eschewed getting locked into Nvidia, paying the high prices state-of-the-art AI technology commands. This penny-smart but pound-foolish approach left the world’s largest consumer electronics company (Apple) and the undisputed cloud leader (AWS) far behind, just when generative AI created massive end-user opportunities they could not adequately address.
Nvidia’s CUDA platform is the de facto standard for training and deploying large language models for generative AI. CUDA offers unmatched performance, developer tooling, and ecosystem support. Companies that build on CUDA — such as Microsoft (in collaboration with OpenAI) and Meta (in partnership with LLaMA) — have been able to scale quickly and deliver cutting-edge AI products.
By contrast, Amazon and Apple chose to go their own way, and Tesla took the Nvidia off-ramp in 2019. Let’s take a look at how each took a different approach and mostly failed.
Apple Maintains Its ABN Strategy (Anybody But Nvidia)
Apple’s generative AI journey has been even more problematic. After unveiling “Apple Intelligence” in 2024, the company’s most anticipated upgrade — a fully LLM-powered Siri — has been delayed until 2026.
Apple has some serious semiconductor bona fides, with its M-class Arm-based chips for desktops and the A-class for mobile. The company is justifiably proud of these efforts. However, Apple initially attempted to leverage AI acceleration, utilizing its own chips, before transitioning to Google TPU-based AI development. Not a bad choice, mind you, but the TPU does not have the performance nor the AI development tool-set that Nvidia Blackwell enjoys. The result? Well, how’s that AI-enhanced Siri and Apple Intelligence working out for you? Yeah, not at all.
To be sure, Apple has significant technical challenges that come with having an installed base and a focus on privacy above all, but not using Nvidia from the start probably cost it more in extra work and time to market than the “expensive” Nvidia infrastructure would have cost.
Siri’s architecture, built over a decade ago, wasn’t designed for generative AI, and retrofitting it has proven more difficult and is taking longer than Apple expected. To make matters worse, Apple’s AI teams have faced internal fragmentation, with some pushing for in-house developed AI models and others advocating partnerships with OpenAI, Perplexity, or Google.
The company also lost key talent to competitors. Ruoming Pang, who led Apple’s foundation models team, left for Meta in 2023. Other researchers followed, citing slow progress and lack of clarity in Apple’s AI strategy.
Amazon AWS Does Offer Nvidia GPUs, but Prefers its Own Silicon
AWS recently paid the price for its slow generative AI sales on Wall Street, which were hindered by Amazon’s hubris and the NIH (Not Invented Here) syndrome. The market share of new generative AI use cases landing on AWS is reportedly lower than its overall cloud share, with Microsoft taking over the lead.
According to IOT-Analytics, Microsoft has a 16% share of new-gen AI case studies, which is comparable to AWS’s share of 37% in 2023. AWS is not losing its first-place share in the overall cloud market, at least not yet. However, for genAI-specific apps and new enterprise AI workloads, Azure and Google are increasingly competitive, and in some cases, are outpacing AWS in genAI-related tools and adoption.
Reducing reliance on Nvidia and lowering costs sounded like a good strategy. So, Amazon’s AWS division, like Google and Microsoft, invested heavily in custom silicon for training and inference, named, of course, Trainium and Inferentia. The latest release, Trainium2, was launched in 2024 and appears to offer impressive specs: up to 83.2 petaflops of FP8 compute and 6 TB of HBM3 memory bandwidth. Amazon even created a 40,000-chip Trainium UltraCluster to support the workloads of generative AI.
But accelerator performance alone doesn’t create AI. You need software, great chip-to-chip networking, and a thriving developer ecosystem. AWS developers found Trainium software more challenging to work with than CUDA, and they reportedly pushed back against management regarding Trainium’s limitations. Management essentially said Shut up and get to work. So, Trainium adoption lagged.
Amazon recognized the need to invest even more in creating the developer ecosystem and launched the Build on Trainium initiative — a $110 million investment in university research. While appealing, this effort came years after Nvidia had firmly cemented its dominance in AI research and development. That is $110 million that could have been better spent on Nvidia hardware and better AI. And that $110 million is in addition to the money AWS spent developing the Trainium and Inferentia chips, which is likely well over $1 billion.
So, Amazon has decided to invest an additional $4 billion in Anthropic, the company behind Claude. Anthropic agreed to use Trainium chips for training its models in return. However, behind the scenes, tensions began to emerge. Engineers at Anthropic reportedly also pushed back against Trainium. Many preferred Nvidia’s stack for its maturity and tooling. Anthropic teams had to rework their CUDA-based pipelines to accommodate Trainium, resulting in delays and performance issues. While Amazon touted the partnership as a breakthrough, it was a compromise — Anthropic needed funding, and Amazon needed a flagship AI partner.
Amazon appears to be changing course lately, deepening its partnership with Anthropic and expanding support for Nvidia GPUs. AWS is building a massive Nvidia cloud infrastructure, Project Ceiba, with over 20,000 Nvidia GPUs. But it is only available to Nvidia engineers for use in developing AI and chips, not for public cloud access.
Now Tesla has seen the Light.
In 2019, Tesla shifted from using Nvidia to its custom FSD Chip for vehicle Autopilot hardware and neural network inference, replacing Nvidia’s Drive PX2 system. And it began a major effort to build its own AI Supercomputer, DOJO, with its in-house chips. Since 2019, Tesla has reportedly spent over $1 billion developing DOJO, along with another $500 million developing a DOJO supercomputer in Buffalo, New York.
Last week, Elon Musk announced on X that he was ending this program and would instead deploy on Nvidia and AMD GPUs. I suspect Tesla will mostly deploy Nvidia this year and see how AMD’s MI400 looks in 2026.
Should Cloud Service Providers Even Build Their Own AI Chips?
Well, first, let’s look at a company that did not. OpenAI has recently reached $12 billion in annualized revenue and surpassed the $700 million weekly active user mark for ChatGPT. And guess what it uses? Yep, Nvidia. Sam Altman does have the gleam of OpenAI chips in his eye, to be sure, but he also realizes that speed, ease of use, and development time matter more to OpenAI than the savings that proprietary chips could provide. At least for now.
Meta has its own MTIA chip, but it is used for internal workloads, like recommendation engines for its Facebook and other properties. Microsoft has its own Maia chips starting with the Maia 100, announced in 2023, used primarily for internal testing and select workloads. The planned successor, Maia 200, is now expected in 2026 due to delays. Maia 200 is designed for data center AI acceleration and inference workloads. We will see if Microsoft learns from Tesla and Apple’s mistakes.
I suspect Google is perhaps the only one to realize a decent return on its AI ASIC (TPU) investments. Still, it has generally failed to attract large outside customers, aside from Apple. But it gets a lot of bang for the buck for internal workloads and training.
My advice to CSPs is this: if you can get Nvidia GPUs, use them. If you have a workload for which you believe they are not ideal and can model a decent ROI, then go for it. Otherwise, save your capital.
The Consequences of Skipping Nvidia Can be Dire
A year in the world of generative AI can mean the difference between success and failure, or at least multi-billion-dollar successes or failures. The hubris of some high-tech companies has cost them billions of dollars, spent needlessly.
Amazon ceded early leadership in cloud AI to Microsoft Azure, which now hosts many of the world’s top models. Apple missed the “AI supercycle” for iPhone upgrades, as consumers saw little reason to buy new devices without meaningful Siri improvements. Tesla has seen the light and is moving fast. All three of these companies now face pressure to catch up — not just in model performance, but in developer mindshare and ecosystem momentum.
Yeah, you can build your own AI chip. But you might regret it.
This article originally appeared on my Substack. Subscribe for free to receive new posts and support my work.