You hear it all the time. Computing power is exploding. AI is getting smarter by the day. Your phone is more powerful than the computers that sent men to the moon. It feels like a runaway train. But if you stop and ask, "How fast is computing power actually growing?" you get a mix of marketing claims, outdated rules of thumb, and genuine scientific debate. The simple answer—the one that matters for your investments and your understanding of the future—isn't simple at all. Growth isn't on a single, predictable track anymore. It's a multi-lane highway, and some lanes are moving much faster than others, while a few old ones are hitting serious traffic. Let's cut through the noise.

The Classic Benchmark: Moore's Law and Its Evolution

For decades, the speed question had a neat answer: Gordon Moore's 1965 observation. He predicted the number of transistors on a chip would double about every two years. For a long time, it did. This became a self-fulfilling prophecy, a roadmap for the entire semiconductor industry. Performance and efficiency rode this wave beautifully.

But here's the non-consensus part everyone glosses over: Moore's Law was always about transistor density, not directly about computing speed or performance. We got lucky for a while because smaller transistors were also faster and used less power. That luck has run into physical walls. As transistors approach the size of atoms, quantum effects and heat dissipation become monstrous problems. The cost of building the factories (fabs) to make these chips is growing exponentially, too. I've talked to engineers who say the "easy" scaling ended over a decade ago. Today, maintaining density growth requires Herculean efforts in materials science and manufacturing—it's no longer a given.

So, if we measure growth purely by transistor count doubling every two years, that pace has slowed. The interval has stretched. The industry still pushes forward, but the old, predictable rhythm is gone. Relying on this single metric to gauge computing power growth is like judging a modern car only by its engine displacement. You'll miss the whole story.

The Multi-Dimensional Engine of Modern Growth

This is where it gets interesting. While the transistor scaling lane slowed, engineers and companies blasted open new ones. Computing power growth didn't stop; it changed direction. Think of it as a shift from making a single engine more powerful to designing a whole new propulsion system.

1. Architectural Innovation: The New Performance King

This is the biggest driver you're not hearing enough about. Companies like Apple, NVIDIA, and AMD are achieving performance leaps not by waiting for the next transistor shrink, but by redesigning the brain of the chip itself.

Apple's M-series chips are a masterclass. By moving away from Intel's generic designs and creating a "system on a chip" (SoC) with unified memory and dedicated blocks for video, AI, and image processing, they delivered a performance-per-watt jump that stunned the industry. I remember benchmarking an M1 Mac against an Intel one I owned. The difference wasn't incremental; it felt like a product from two different eras. That leap came from architecture, not just a die shrink.

2. The Specialization Surge: GPUs, TPUs, and NPUs

General-purpose CPUs are becoming the jack-of-all-trades in a world of masters. For specific workloads, specialized processors are growing in capability at a blistering, non-linear rate.

• GPUs for AI: NVIDIA's hardware for AI training has seen performance multiply by orders of magnitude in a few years, driven by architecture (tensor cores), better software (CUDA), and stacking memory (HBM). The growth curve for this type of computing power is almost vertical.
• Domain-Specific Chips: Google's TPUs, chips for cryptocurrency mining (historically), and modern networking chips are all examples. They do one thing exceptionally well, and their growth in that specific task dwarfs general CPU progress.

3. The Software and Algorithm Layer

This is the silent multiplier. A better algorithm can make a 10-year-old computer solve a problem faster than a new one running a brute-force approach. Machine learning frameworks like TensorFlow and PyTorch, along with compiler improvements, are constantly extracting more performance from the same silicon. Growth here is hard to quantify in GHz, but it's real and significant.

Measuring the Speed: It's Not Just About Transistors

So how do we measure this multi-faceted growth? You have to look at different benchmarks for different jobs.

Look at the TOP500 list of supercomputers. The growth in FLOPS has been remarkably steady, but that's because it's achieved by throwing more parallel processors (often GPUs) at the problem, not just making each one faster. It's a different kind of growth—broad, not deep.

The takeaway? There is no single speed. Computing power for AI tasks is growing explosively. For everyday general tasks, it's growing steadily but unspectacularly. The average is meaningless. You must look at the segment.

The Future Trajectory: What's Next for Computing Power?

The next decade will be defined by diversification and hitting new walls.

Heterogeneous Integration: Growth will come from stacking different specialized chips (CPUs, GPUs, memory) in a single package, like a high-tech lasagna. This improves speed and reduces power, but adds design complexity. Intel's Foveros and AMD's 3D V-Cache are early examples.

The Memory Wall: This is the next big bottleneck. Processors are becoming so fast that waiting for data from memory is a huge drag. Progress in memory bandwidth (like HBM) and new architectures (processing-in-memory) will be critical to sustaining growth. If this isn't solved, faster cores will just sit idle more often.

Post-Silicon Explorations: Quantum computing (for specific problems), photonic chips, and neuromorphic computing are wildcards. They won't replace silicon for general computing soon, but they represent potential new curves of growth for niche, high-value applications. Don't believe the hype that they're around the corner for your laptop, but do watch the research.

Computing Growth and Your Wallet: Investment Implications

If you're thinking about this from an investment angle (tech stocks, sector ETFs), the shifting growth dynamics are crucial. Investing in "computing power" is no longer a monolithic bet.

Betting on the Specialization Lane: This means companies leading in AI hardware (NVIDIA, AMD to a degree), designers of domain-specific chips, and firms with superior architectural skills (Apple, though it's not a pure-play). Their growth trajectory is steeper, but valuations often reflect that. The risk is technological disruption or a slowdown in AI spending.

Betting on the Foundry/Manufacturing Lane: Companies like TSMC and ASML. They enable everyone else's growth. It's a more stable, oligopolistic bet, but tied to the massive capital expenditure cycles of the industry. Growth here is solid but cyclical.

The Caution: Companies reliant solely on selling traditional, general-purpose CPUs without a compelling roadmap for AI or efficiency are in a tougher spot. Their growth lane is the slower, more congested one. They must reinvent themselves.

My approach has been to look for companies that control the entire stack—hardware, architecture, software. That integration is where they can squeeze out performance gains others can't, creating a moat. It's also why I'm wary of companies that just market "more cores" or "higher GHz" without context.

So, how fast is computing power growing? It depends. The age of one simple answer is over. We're in an era of divergent growth curves. For AI and specialized tasks, it's breathtakingly fast, driven by new silicon architectures. For general-purpose computing, it's modest and increasingly focused on efficiency. This complexity is what makes the field both challenging and endlessly fascinating. It also means that understanding these nuances is the key to making smart decisions, whether you're building a product, planning a strategy, or deciding where to invest your money. Ignore the headline hype. Look at the benchmarks that matter for the task at hand. That's where you'll find the real speed.