Google just shipped Gemini 2.0 with enhanced multimodal reasoning. AMD shipped the MI300X with 180GB of HBM3 memory. Apple shipped the M4 with a Neural Engine that runs ML workloads locally. Three companies, three different bets on where the AI stack consolidates. I think only one of these bets survives contact with the market, and it is not the one the headlines are calling the winner.

What Google Is Actually Betting On

Gemini 2.0 is not just a model update. It is Google’s thesis that multimodal reasoning is the defensive moat against OpenAI’s text dominance and Anthropic’s safety positioning. The model handles image and text processing in the same inference pass, which matters because most real-world queries are not pure text. “Show me hotels near this landmark” requires vision and search and maps context simultaneously.

Per The Verge’s reporting, the reasoning improvements are measurable across benchmarks that combine modalities. Google is not claiming AGI. It is claiming that the models that win enterprise adoption will be the ones that handle messy, real-world queries where text alone is not enough.

The counter-narrative writes itself: OpenAI’s GPT-4 already does vision, Anthropic’s Claude 3 Opus already does documents, what is Google’s actual edge here? The edge is distribution. Gemini 2.0 ships in Search, in Workspace, in Android. The model that reaches a billion users daily does not need to be 10% better. It needs to be 2% better in the places those users already are.

What AMD Is Actually Betting On

The MI300X is AMD’s bet that memory bandwidth matters more than CUDA lock-in. 180GB of HBM3 memory is not a spec bump. It is an architectural statement. Training runs hit memory limits before they hit compute limits on most frontier models. Per TechPowerUp’s specs, the MI300X delivers 5.2 TB/s of memory bandwidth, against Nvidia’s H100 at 3.35 TB/s.

The thesis: if you can fit the entire model in HBM and eliminate DRAM swapping, you cut training time by 30-40%. That savings translates directly to cost savings at hyperscale. A cluster of MI300X chips running PyTorch does not care about CUDA if the framework abstraction layer is good enough and the economics are 25% cheaper per training run.

The risk: CUDA is not just a programming model. It is two decades of optimized libraries, debugged kernels, and institutional knowledge. AMD’s ROCm software stack has improved dramatically, but “improved dramatically” and “production-ready at Nvidia parity” are not the same thing. Every AI engineer I know can write CUDA. Half of them have heard of ROCm. That gap is AMD’s marketing problem and its technical opportunity simultaneously.

What Apple Is Actually Betting On

As a declared Snapdragon Insider who covers the on-device AI space closely, I want to name the bias before making the argument: I think Apple’s bet is the most defensible, but I have a professional and editorial interest in the on-device AI thesis.

The M4’s Neural Engine is Apple’s bet that on-device AI is now a business decision, not a feature. 16GB of unified memory is not about performance. It is about independence. A model that runs locally cannot be rate-limited, cannot be revoked by export controls, and cannot leak your data to a third-party inference endpoint.

Per 9to5Mac’s coverage, the M4’s Neural Engine delivers up to 38 TOPS, double the M3’s 18 TOPS. For context, the Snapdragon X Elite’s Hexagon NPU hits 45 TOPS, and Snapdragon 8 Gen 3 hits 73 TOPS in burst mode. Apple is competitive, not class-leading. But Apple’s advantage is the software stack: Core ML integration, OS-level inference pipeline, and the tightest hardware-software coupling in the consumer device market.

The counter-bet is obvious: cloud models are smarter. GPT-4 runs on a cluster. Gemini 2.0 runs on TPUs. The M4 runs on 16GB of RAM. There is no universe where local models match cloud models on raw capability. Apple’s thesis is that 80% capability with 100% privacy beats 100% capability with 0% privacy for the workflows that matter on a consumer device.

Why Only One of These Bets Wins

The market does not have room for three parallel AI stacks. Developers will standardize on one inference pattern, one memory architecture, one deployment model. The winner will be the company that solves the actual bottleneck, not the company with the best benchmark numbers.

Google’s bet wins if multimodal reasoning becomes table stakes and distribution is the moat. AMD’s bet wins if memory bandwidth is the limiting factor and CUDA lock-in breaks under economic pressure. Apple’s bet wins if privacy regulation and sovereign AI concerns push inference back to the edge.

I think Apple’s bet is the most defensible, but not for the reason most people assume. It is not about privacy. It is about control. Every time a developer builds for on-device inference, they are building something that cannot be shut down by a third party. Every time a user runs a model locally, they are running something that does not require network access, API keys, or rate limits.

The AI economy right now is built on rented intelligence. You pay per token. You accept rate limits. You trust that the endpoint stays online. When third-party API access gets restricted, every application built on top of it stops working instantly. Nobody can revoke your local weights.

Google’s bet requires that multimodal queries stay inside Google’s ecosystem. AMD’s bet requires that data centers prioritize cost over compatibility. Apple’s bet requires nothing except that the user owns the hardware. That asymmetry is why I think Apple’s architecture survives even if Google’s model is smarter and AMD’s chip is faster.

I laid out the full stack split argument in June, before Gemini 2.0 dropped. Gemini 2.0 confirms the cloud thesis is getting stronger in absolute terms. It does not change the sovereignty thesis. The question is not which company ships the best AI. The question is which architecture developers build for when they realize rented intelligence has a kill switch.