On specialization and the pursuit of record performance

I was in Tokyo this week for the STAC Summit. By coincidence, the World Athletics Championships were happening at the same time. One evening I was at the Japan National Stadium watching sprinters and distance runners push human limits. The next morning I was at our Summit, where firms shared the latest advances in high-performance technology. At first those felt like two different worlds. But the more I thought about it, the more I realized they were the same world, playing out on different stages.

Both were about specialization.

When you see elite athletes up close, you realize how narrow their optimization really is. A 100-meter sprinter and a marathoner are technically both “runners,” but their bodies, training, and equipment are radically different. Even their shoes are unrecognizable as belonging to the same sport. A sprinter’s ASICS spikes are engineered for explosive starts and raw acceleration; a marathoner’s ASICS shoes are engineered for efficiency and endurance. The only thing they share is the logo.

This is the pattern of high performance: narrow the problem, then optimize relentlessly.

The ASICS name itself comes from the Latin phrase Anima Sana in Corpore Sano — “a sound mind in a sound body.” The phrase captures balance: to perform at the limit, every part of the system — muscles, lungs, equipment — must be aligned toward a single objective. That alignment is what turns potential into record-breaking performance.

The same happens in data centers. CPUs are the generalists of computing. They can run anything, but they carry overhead: complex instruction sets, layers of caches, branch prediction, power management. GPUs are optimized for a narrower class of workloads, namely parallel operations, but are still broad in terms of the applications they support.

Then there are ASICs — Application-Specific Integrated Circuits — designed to execute one workload with extreme efficiency. They give up flexibility, but in exchange they can deliver determinism and speed that general-purpose processors cannot.

CPUs are like marathoners: versatile, steady, and able to handle almost anything you throw at them — but never the absolute fastest for any single task. GPUs are closer to sprinters: explosive in their power, designed for one kind of race, but less suited to the long, meandering workloads of general computing. ASICs are like the sprinter’s spikes — engineered so precisely for a single race that they look useless outside of it. But when you’re in that lane, they can deliver a performance no generalist can touch.

At STAC, ASICs are part of our discussions at every Summit. And they fit squarely into a longer story our benchmarks document: the steady march of specialization in finance.

Fifteen years ago, trading workloads ran almost exclusively on CPUs. Then came the FPGA wave — firms realized they could hard-wire critical paths like order book updates and risk checks. Around the same time, GPUs entered the picture for research workloads, unlocking levels of parallel simulation that CPUs couldn’t approach. Now, trading firms are looking at ASICs and other accelerators as the next phase: chips designed not for general parallelism but for specific, repeated algorithms at massive scale.

The trajectory is unmistakable. Each generation of specialization has allowed firms to do something they couldn’t do before.

That’s why records matter. In athletics, when a runner breaks a world record, it doesn’t just prove they are the best in that moment. It redefines what is possible. Everyone else in the sport has to recalibrate their expectations. Coaches design new training regimes. Shoe companies design new materials. Young athletes grow up knowing the ceiling is higher than their predecessors believed.

The same is true in finance. When a new system posts breakthrough performance on a STAC benchmark, it isn’t just about bragging rights. It changes how firms think about the economics of trading and risk. If latency can be reduced from hundreds of microseconds to tens, entirely new categories of machine learning models become usable in live trading systems. If backtesting throughput increases by an order of magnitude, research teams can explore strategies that were previously impractical. If time-series queries drop from hours to seconds, risk managers can interrogate exposures intraday instead of overnight.

In both cases, a record resets the industry’s collective imagination.

Finance is ultimately about making decisions under uncertainty. There is a concept in portfolio theory called the efficient frontier: the set of portfolios that achieve the highest return for a given level of risk. What’s often overlooked is that the efficient frontier is not fixed. It shifts with technology.

When benchmarks demonstrate that you can run more complex simulations, or process more market data in real time, the entire frontier moves outward. What once required unacceptable latency or cost is suddenly feasible. The economics of trading, risk, and research change because the technological boundary has moved.

That’s the significance of records. They expand the set of possible strategies, just as a new world record in athletics expands the set of possible performances.

Benchmarks are the mechanism by which the industry measures these shifts. Without them, every vendor claim would blur into noise. With them, we can see the difference between incremental improvement and genuine breakthroughs.

Athletics has always understood this. Records are not anecdotes; they are standardized measures under strict rules. That’s why they inspire confidence. STAC was built on the same principle. By providing industry-designed workloads and independently audited results, we make it possible to distinguish hype from real progress.

This matters not just for engineers, but for the industry as a whole. Investors, traders, risk managers — they all need to know where the frontier really is, because it tells them what decisions they can make in practice, not just in theory.

There’s a paradox here. Specialization looks like it reduces options. A sprinter cannot win marathons. An ASIC cannot run general workloads. A trading system designed for extreme determinism will not excel at throughput.

But in practice, specialization creates options at the frontier. It allows a new record to be set, and that record shifts the frontier outward for everyone else.

The lesson from Tokyo is that progress comes not from trying to be good at everything, but from narrowing down and optimizing relentlessly for one thing. This principle applies as much to startups and research as it does to athletes and silicon. Focus looks like a constraint, but it is actually the precondition for breakthroughs.

So what I took away from Tokyo wasn’t just that I happened to see two sets of records fall in two days. It was that they were telling the same story.

On the track, records fell because athletes and engineers focused obsessively on the smallest details: stride length, shoe materials, wind resistance. In the data center, records fell because firms focused obsessively on system design: chip architecture, interconnect fabrics, software tuning.

Both show us that excellence is not a general pursuit. It’s the product of specialization, discipline, and measurement.