The Hardware Paradox: Why the "Old" Tech Keeps Winning
A Tanikota Market Diary — Engineering, Tennis & Buffett
The Opening Hook
When we look at the AI gold rush, the instinct is to bet on the "shiny new pickaxe." Startups are building purpose-built silicon that promises to run rings around the incumbents on paper.
But in my years of tracking tech transitions, I've learned one hard truth: Versatility and Infrastructure outlast Raw Speed. Let me walk you through the three pillars of why I believe the status quo isn't going anywhere soon — drawn from aviation, my love for tennis, and the wisdom of Omaha.
1. The "Fossil-Fuel" Reality of Computing
Think about the jet engine. It's noisy, fuel-hungry, and mechanically complex. Yet, it dominates aviation because it runs on any fuel, operates in any weather, and fits every airframe.
General-purpose GPUs are the jet engines of AI.
Specialized alternatives (hardwired chips) might boast incredible inference speeds, but they lack the flexibility to handle new AI architectures. The industry is still figuring out if Transformers are the final stop, or just a layover. If you hardwire your silicon for today's math, you risk becoming obsolete tomorrow. The incumbent remains the safe, reliable baseline that the entire industry builds its runways around.
2. The Tennis Analogy ðŸ
I spend a lot of time on the tennis circuit, and I see the same dynamic playing out between traditional tennis and newer paddle sports.
Paddle sports have massive sponsorships, state-of-the-art racquet technology, and deep pockets. They are innovative and fun. But Tennis is Tennis.
It owns the four Grand Slams — Wimbledon, Roland Garros, the US and Australian Opens. It owns the ATP and WTA tours. These are the canonical events.
In AI, the "Grand Slams" are the training clusters, the enterprise data centers, and the CUDA software stacks. You can have better "paddle" tech for inference, but if you aren't playing on the main tour — if you aren't training your models where the champions train — you aren't shaping the game. You are just a very talented exhibition player.
3. The Buffett Epilogue: Valuing the Moat, Not the Speed
This brings me to the Oracle of Omaha. Warren Buffett doesn't buy the fastest-moving stocks; he buys the ones that are hardest to move away from.
ð¹ The Moat: The market leader isn't just selling chips; they are selling an ecosystem. Migrating away from this software stack is a multi-year, multi-million-dollar headache for companies. That is a moat wide enough to sail a battleship through.
ð¹ The Ark vs. The Sail: Buffett famously said, "Predicting the rain doesn't count; building the ark does." The general-purpose chip is the ark. It weathers the storm of changing AI models (Mamba, Transformers, or whatever comes next). The specialized chips are faster sails — they catch today's wind perfectly, but they sink if the wind shifts.
ð¹ The Circle of Competence: I understand railroads; they move cargo reliably. I understand insurance; it is actuarially predictable. General-purpose GPUs are like railroads — they have the right-of-way, the installed tracks, and the long-term contracts. Specialized startups are building hyperloops: exciting, but they haven't broken ground yet.
"In the short run, the market votes for the best benchmark.
In the long run, it weighs the best infrastructure."
The Coexistence Conclusion
To be clear: I am not saying the specialists will disappear.
Just as electric drives are finding their place in specific transport sectors, and paddle sports are growing their dedicated fanbase, these new AI chips will capture high-volume, standardized inference workloads.
However, for the cutting-edge research, the mission-critical enterprise deployments, and the training of the next billion-parameter models, the industry will stick with the generalist backbone.
And right now, the scales are heavily tipped toward the system that owns the calendar, the tracks, and the Grand Slams.
ð Glossary of Terms
Plain-English translations for the non-tech crowd
GPU (Graphics Processing Unit)
Plain English: A specialized computer chip originally designed to render video game graphics. It turns out this chip is also brilliant at doing many math problems at the same time — which is exactly what AI needs. Think of it as a team of 10,000 calculators working in parallel.
ASIC (Application-Specific Integrated Circuit)
Plain English: A chip that is hardwired to do just one job — and do it incredibly fast. Like a dedicated ice cream maker that only makes vanilla, but makes it perfectly. The catch? If everyone suddenly wants chocolate, you're stuck.
Inference vs. Training
Plain English: Training is like sending a student to school for years — feeding them billions of examples until they learn. Inference is the final exam — when you actually ask the AI a question and it gives you an answer. Training is heavy lifting; inference is the daily usage.
CUDA (Compute Unified Device Architecture)
Plain English: A software language and toolkit created by NVIDIA that lets programmers talk to GPUs. It's like the operating system for AI chips. Millions of developers have learned this language. Switching to a different chip means learning a whole new language — expensive and time-consuming.
Transformer Model
Plain English: The current rockstar architecture behind AI like ChatGPT. It works by paying "attention" to the most important parts of the input. It's what made modern AI explode. But tech moves fast — tomorrow's AI might use something completely different.
Moat (Business)
Plain English: Warren Buffett's favorite word. A competitive advantage that protects a company from rivals. Like a castle moat filled with crocodiles — it makes it really hard for competitors to attack. For the leading chip maker, the moat is the millions of developers who don't want to learn a new system.
Ecosystem
Plain English: Not nature — it's the entire network of software, tools, developers, and training that grows around a product. Apple's iPhone has an ecosystem (apps, developers, accessories). NVIDIA has an AI ecosystem. It's hard to beat because everyone is already inside it.
Foundry / Supply Chain
Plain English: The factories that make computer chips (like TSMC). They are incredibly expensive and have limited space. The big chip company buys most of that factory capacity years in advance. It's like booking all the best tables at a restaurant — startups can't get in.
Disclaimer: This is a personal market diary and thought experiment. It does not constitute financial advice. Always do your own research.