For years, the crypto industry has thrown around big numbers without much context. Bitcoin’s hashrate climbs to 800 exahashes, then a zettahash, and most people nod along without any intuitive sense of what those figures actually mean. At the Proof of Talk summit in Paris this week, Bittensor co-founder and Crucible Labs partner Ala Shaabana offered a comparison designed to make the scale visceral: Bitcoin’s network now operates at more than 600,000 times the combined computing power of the world’s top 100 supercomputers.
That figure is not a typo. It is not rounded up for dramatic effect. And Shaabana’s point in citing it was not simply to marvel at Bitcoin’s growth. His argument is that the same incentive architecture that turned a 2009 whitepaper into the most powerful computing engine on Earth can be repurposed for artificial intelligence, and that this shift is already underway.
The Math Behind 600,000x
Supercomputer rankings are typically measured in FLOPS (floating-point operations per second), while Bitcoin’s hashrate is measured in SHA-256 hashes per second. These are not directly comparable units, which makes Shaabana’s claim somewhat dependent on how you frame the conversion. Still, the spirit of the comparison holds up under scrutiny.
The world’s fastest supercomputer, the Frontier system at Oak Ridge National Laboratory, achieves roughly 1.2 exaFLOPS at peak performance. The combined power of the top 100 supercomputers on the TOP500 list sits somewhere around 5 to 6 exaFLOPS. Bitcoin’s network, meanwhile, recently crossed 700 exahashes per second.
When you translate SHA-256 hashing into equivalent computational work, the ratio balloons into the hundreds of thousands. The precise multiplier depends on your methodology, but the order of magnitude is not in dispute: Bitcoin’s distributed mining network represents a concentration of hardware and electricity that no government or corporation has ever matched.
“We all know that Bitcoin really dwarfs the top 100 supercomputers,” Shaabana told the Paris audience. “Does anybody know, in comparison, what the hash rate is? It’s over 600,000 times the power of really what these supercomputers can do. And that’s just, really, it’s Bitcoin.”
The qualifier at the end is the key. Shaabana co-founded Bittensor, a protocol designed to redirect that same coordination mechanism toward AI. His pitch is that if economic incentives can summon 600,000 supercomputers’ worth of hashing power for a single network, they can do the same for machine learning, data storage, and inference.
How Bittensor Borrows Bitcoin’s Playbook
Bittensor is often described as “Bitcoin for AI,” and the label is more literal than most marketing analogies. The protocol shares Bitcoin’s core economic constraints: a hard cap of 21 million tokens (TAO), halvings hardcoded into predetermined blocks, no pre-mine, and no venture capital round. These design choices are not aesthetic. They are meant to replicate the economic credibility that gave Bitcoin staying power.
The difference is what miners are optimizing for. In Bitcoin, miners compete to solve SHA-256 hash puzzles, and the fastest hardware wins. In Bittensor, miners compete to perform useful AI tasks, from training language models to validating inference outputs to storing data, and the most accurate or efficient work wins.
The network is organized into 128 specialized subnets, each with its own goal and reward structure. A subnet focused on text generation might reward miners for producing coherent responses. A subnet focused on storage might reward miners for data availability and retrieval speed. The design allows the network to coordinate global hardware toward specific problems without a central administrator deciding what matters.
“Show me the subnet, and I’ll tell you what the miners are optimizing for,” Shaabana said, adapting a famous market quote. The implication is that incentive design is the only lever that matters. Get the rewards right, and the hardware and talent follow.
This is a direct inversion of how most AI infrastructure is built today. OpenAI, Google, and Microsoft operate massive centralized data centers, competing for GPU allocations and negotiating power contracts. Bittensor’s bet is that a decentralized network, properly incentivized, can outcompete those monopolies on cost and scale.
Why This Matters for Bitcoin’s Mining Industry
Bitcoin miners have spent years building out infrastructure, from ASIC manufacturing pipelines to substation relationships to cooling systems, that is increasingly specialized for one purpose: hashing SHA-256. As block rewards decline through successive halvings, the question of what else that infrastructure can do becomes more pressing.
Bittensor does not directly use Bitcoin’s mining hardware. SHA-256 ASICs are useless for AI workloads, which require GPUs or specialized AI accelerators. But the economic and coordination lessons transfer. Mining pools that have mastered distributed incentive management, such as the seven pools controlling 75% of Bitcoin’s hashrate that recently backed the Stratum V2 upgrade, are well-positioned to understand how decentralized work coordination functions at scale.
The broader point is that Bitcoin proved something previously considered impossible: that a permissionless network of self-interested participants, with no central authority, could sustain the largest coordinated computing effort in human history. That proof of concept is now being stress-tested on different problems.
The Trust Deficit Driving Decentralization
Shaabana’s Paris presentation did not stop at technical arguments. He tied the decentralized computing thesis to a broader shift in how people relate to institutions.
“The long-term bull case is no longer primarily technological,” he said. “It is driven by debt, liquidity, and declining trust in traditional sovereign systems. Subnets really create markets. Intelligence really is no longer locked behind issues of organization; signals will define the truth, and performance is really rewarded.”
This is a more speculative claim, but it resonates with observable trends. Trust in governments, central banks, and large corporations has declined across most developed economies over the past decade. The appeal of crypto networks, originally, was that they offered financial infrastructure that did not require trusting a counterparty. Bittensor extends that logic to AI: if you do not trust OpenAI or Google to steward general-purpose intelligence, perhaps a network where performance is rewarded and verified on-chain is more appealing.
Whether this framing holds up depends on whether decentralized AI networks can actually match or exceed centralized competitors on output quality. That remains unproven at scale. But the argument from capital flows is harder to dismiss. Bittensor’s TAO token has a market cap north of $3 billion, and the protocol has attracted meaningful mining participation despite being less than three years old.
Practical Implications for AI Cost and Access
One reason Shaabana’s thesis matters is cost. Training large language models at the frontier requires hundreds of millions of dollars in compute. OpenAI’s rumored training runs for GPT-5 class models involve clusters of tens of thousands of H100 GPUs, and the scarcity of those chips has created a bidding war among well-funded labs.
Decentralized networks offer a different path. By aggregating unused or underutilized hardware from around the world, protocols like Bittensor can potentially offer AI inference and training at lower cost than centralized providers. The Paris summit also featured Titan Network, another decentralized compute project, which claims to offer AI infrastructure at up to 75% lower cost than traditional cloud providers. Titan says it has signed tech giants like Tencent and Alibaba as clients.
These claims deserve scrutiny. Centralized providers benefit from colocation efficiencies, predictable latency, and economies of scale that distributed networks struggle to replicate. But for certain workloads, particularly those that are latency-tolerant or embarrassingly parallel, decentralized compute may genuinely offer an economic advantage.
The subnet model also enables specialization that monolithic cloud providers cannot easily match. A subnet focused exclusively on protein folding, for example, can attract hardware and talent optimized for that problem, without the overhead of a general-purpose cloud stack. This is analogous to how Bitcoin mining became progressively more specialized over time, moving from CPUs to GPUs to FPGAs to purpose-built ASICs.
What 600,000x Actually Means for Network Security
Shaabana’s 600,000x figure also carries implications for Bitcoin’s own security model. The energy and hardware expended on Bitcoin mining is not wasted; it is the cost of attack resistance. A network that commands 600,000 times the combined power of the world’s top supercomputers is, by definition, extraordinarily expensive to attack.
Critics have long argued that Bitcoin’s energy consumption is indefensible. Shaabana’s framing inverts that critique: the energy expenditure is the point. It represents a coordination achievement that no alternative mechanism, including proof-of-stake, has replicated at scale. Whether you find that tradeoff acceptable depends on how much value you assign to censorship-resistant money, but the raw computational fact is not in dispute.
For Bittensor, the question is whether a similar level of energy and hardware commitment can be sustained for AI workloads. Mining TAO requires GPUs rather than ASICs, and GPU economics are different, more volatile, and more dependent on semiconductor supply chains controlled by a handful of companies. Nvidia’s market position, in particular, creates a bottleneck that did not exist in Bitcoin’s early years.
Still, the subnet model allows for some flexibility. Different subnets can target different hardware profiles, from consumer GPUs to specialized inference chips to storage nodes. This modularity may help the network adapt to shifting hardware availability more gracefully than a monolithic protocol could.
The Skeptic’s Case
Not everyone in the Paris audience was convinced. Decentralized AI networks face challenges that Bitcoin did not. Hashing is a simple, easily verifiable operation: either your hash meets the difficulty target or it does not. AI tasks are fuzzier. How do you verify that a language model response is “good”? How do you prevent miners from gaming evaluation metrics? How do you handle the coordination overhead of 128 separate subnets, each with its own incentive structure?
These are real problems, and Bittensor’s solutions remain works in progress. The protocol relies on a combination of validator-based scoring and economic staking to prevent manipulation, but edge cases abound. A subnet focused on subjective outputs, like creative writing or image generation, is inherently harder to evaluate than one focused on objective benchmarks.
There is also the question of whether decentralized AI can keep pace with centralized labs on frontier capabilities. OpenAI, Google DeepMind, and Anthropic have access to the best researchers, the most data, and the most capital. A decentralized network might offer cost advantages for commoditized inference, but the bleeding edge of model development may remain centralized for years or decades.
Shaabana’s response is that the goal is not to beat OpenAI at its own game, at least not immediately. The goal is to create an alternative infrastructure layer that is not subject to corporate gatekeeping. If a single company decides that certain AI capabilities are too dangerous to release, or demands too high a price for access, a decentralized network offers a release valve.
Where This Goes From Here
The Paris summit marked a moment where two narratives, Bitcoin as a monetary network and decentralized AI as the next frontier, were explicitly linked by someone building in both spaces. Shaabana’s 600,000x comparison was designed to make that link intuitive.
Bitcoin’s network did not become the world’s largest computing system because Satoshi Nakamoto asked nicely. It grew because miners were paid to compete. Bittensor’s bet is that the same dynamic can be replicated for intelligence, and that the result will be an AI infrastructure layer that no government or corporation can shut down.
Whether that bet pays off depends on execution. The subnet model is elegant in theory but messy in practice. Incentive design is hard, and getting it wrong can lead to perverse outcomes. But the existence of Bitcoin, operating for 17 years with no downtime and no central administrator, is proof that the approach can work.
For investors and builders, the implication is that decentralized compute is no longer a fringe idea. It is a thesis with billions of dollars behind it and a live network processing real workloads. The 600,000x number is memorable precisely because it is outlandish, and outlandish numbers have a way of demanding attention.
