Huawei Chip Scientist Warns AI Hardware Scaling Is Nearing a “Avalanche” Point: What It Means for Crypto Infrastructure

The next bottleneck in crypto may not be block space, liquidity, or regulation. It may be silicon.

A recent warning from Liao Heng, Huawei’s chief semiconductor scientist and one of the key figures behind the Ascend AI chip effort, has sparked discussion far beyond the AI industry. His core message is simple but consequential: the dominant path of scaling AI compute by stacking more processors and more high-bandwidth memory cannot continue indefinitely. At some point, physical limits become economic limits, and economic limits become system-wide constraints.

For the blockchain and cryptocurrency industry, this is not just an AI hardware story. It is a preview of the infrastructure pressures that will shape zero-knowledge proof generation, decentralized compute markets, AI agents on-chain, validator operations, data availability, and the next generation of crypto security.

The End of “Just Add More GPUs”

For the past decade, the AI industry has relied heavily on brute-force scaling: more chips, larger clusters, faster interconnects, and more memory bandwidth. NVIDIA’s GPU ecosystem became the default foundation for large-scale AI training, inference, and increasingly, adjacent workloads such as cryptographic proving and simulation. The company’s own materials emphasize accelerated computing as a full-stack approach across hardware, networking, software, and developer tooling via the NVIDIA data center platform.

Liao’s warning challenges the assumption that this scaling strategy can continue smoothly. His argument is not that AI compute demand will disappear. Rather, it is that repeatedly increasing chip count and memory capacity eventually runs into constraints: power density, heat dissipation, packaging complexity, interconnect latency, supply chain concentration, and cost.

This matters for crypto because many emerging blockchain workloads are becoming more compute-intensive, not less.

Zero-knowledge rollups need proof generation. Fully homomorphic encryption experiments need specialized acceleration. On-chain AI applications depend on inference layers. Decentralized physical infrastructure networks require verification. Restaking and shared security designs introduce more monitoring, simulation, and automation. In short, crypto is moving from a transaction-centric era into a computation-centric era.

If the AI hardware curve becomes more expensive or less predictable, crypto infrastructure builders will feel it.

Why AI Chip Limits Matter to Blockchain

Blockchain systems have always been shaped by hardware realities.

Bitcoin mining was transformed by ASICs. Ethereum’s transition to proof of stake reduced energy-intensive mining, a shift documented by the Ethereum Foundation’s energy consumption research. Solana’s high-throughput design assumes validators can maintain demanding hardware and networking standards. Zero-knowledge networks depend on provers that often require powerful CPUs, GPUs, FPGAs, or custom accelerators.

The industry sometimes describes crypto as “software eating finance,” but in practice, every blockchain depends on physical infrastructure:

• Validator hardware and network latency
• Data centers and cloud providers
• Secure key storage devices
• Proof generation machines
• RPC infrastructure
• Indexers and archive nodes
• AI inference endpoints for automated agents
• Energy supply and geographic distribution

A slowdown in traditional chip scaling could push blockchain teams to rethink architecture in the same way AI researchers are being forced to rethink model design.

Instead of assuming infinite compute, protocols may need to optimize for efficiency, specialization, and locality.

The “18-Story Pagoda” View of AI Infrastructure

Liao reportedly described the AI value chain as an “18-story pagoda,” contrasting with Jensen Huang’s well-known layered framing of the AI stack. The metaphor is useful for crypto as well.

A modern blockchain application is not just a smart contract. It sits on top of many interdependent layers:

1. Chip design and fabrication
2. Memory and packaging
3. Data centers and power supply
4. Operating systems and drivers
5. Cryptographic libraries
6. Virtual machines
7. Consensus mechanisms
8. Execution clients
9. Data availability layers
10. Oracles
11. Indexers
12. Wallets
13. Front ends
14. Security monitoring
15. Governance systems
16. Compliance tooling
17. AI automation
18. User trust and custody

The lesson is that competitiveness does not come from one layer alone. It comes from coordination across the stack.

This is especially relevant to zero-knowledge infrastructure. A faster proving system is not only a better algorithm; it also depends on memory access patterns, parallelization, compiler design, circuit architecture, and hardware acceleration. The same is true for decentralized AI networks, where token incentives alone cannot overcome inefficient compute routing or unreliable inference performance.

Crypto protocols that treat infrastructure as a full-stack problem will likely be better positioned than those that assume general-purpose cloud compute will always be cheap and abundant.

Tau Scaling Law and the Shift From Miniaturization to Data Movement

One of the most interesting ideas associated with Huawei’s chip strategy is Tau Scaling Law. The concept, as described by Liao, is that once the traditional benefits of shrinking transistors begin to weaken, system designers should focus more aggressively on improving data transfer between components.

This is highly relevant to blockchain.

Many crypto workloads are not limited only by raw computation. They are limited by memory movement, communication overhead, and coordination costs. For example:

• ZK proof generation often involves large intermediate data structures.
• Validator performance can depend on state access and network propagation.
• Rollup sequencers must coordinate ordering, execution, and settlement.
• AI agents interacting with blockchain systems need fast inference plus reliable signing flows.
• Decentralized storage and compute networks must move data efficiently across untrusted environments.

In other words, the bottleneck is often not “can we compute this?” but “can we move, verify, and coordinate the required data efficiently enough?”

That is why chip-level thinking increasingly overlaps with protocol-level thinking. A blockchain scaling roadmap that ignores hardware realities may look elegant in theory but struggle in production.

The DeepSeek Lesson: Efficiency as a Strategic Advantage

Liao also praised DeepSeek founder Liang Wenfeng, arguing that the company’s breakthrough was not merely about using fewer resources, but about designing model architectures that made better use of limited compute.

The broader lesson for crypto is clear: efficiency can be a moat.

In 2025, DeepSeek became a global reference point for cost-conscious AI development, as its research and models attracted attention through the company’s publications on arXiv and open model releases. The significance was not simply that a model could be cheaper. It was that architectural innovation could narrow the gap against competitors with larger hardware budgets.

Crypto has a similar dynamic.

A protocol with fewer resources can still compete if it reduces verification costs, compresses state more effectively, improves prover efficiency, or designs incentives that minimize waste. This is already visible in the rollup ecosystem, where teams are experimenting with more efficient proof systems, alternative virtual machines, and modular data availability designs.

The next cycle may reward protocols that do more with less.

Decentralized Compute Will Face a Reality Check

The chip scaling debate also affects one of the most popular crypto narratives: decentralized compute.

The idea is attractive. Instead of relying on a few hyperscale cloud providers, networks can coordinate idle GPUs, specialized hardware, and distributed data centers through token incentives. This could support AI inference, rendering, scientific computing, or cryptographic proving.

But hardware scarcity cannot be solved by tokenomics alone.

A decentralized compute network must answer hard questions:

• Are the available chips suitable for the workload?
• Can performance be verified without excessive overhead?
• Is data transfer fast enough?
• Can users rely on uptime and predictable pricing?
• How does the network handle privacy-sensitive workloads?
• Are incentives aligned for long-term infrastructure investment?

If high-end chips become more constrained, decentralized compute markets may become more valuable, but also more demanding. Users will not just want access to “a GPU.” They will want verified performance, secure execution, predictable latency, and transparent pricing.

This is where blockchain can add value: settlement, reputation, escrow, access control, and auditable usage records. But the underlying hardware layer still matters.

ZK, AI, and the Coming Demand for Specialized Hardware

Zero-knowledge technology is one of the clearest areas where chip constraints and blockchain strategy intersect.

ZK proofs promise scalable verification: a blockchain can verify a compact proof instead of re-executing a large computation. This is central to many rollup designs and to Ethereum’s long-term scaling roadmap, including the broader move toward rollup-centric execution described in Ethereum research discussions such as Ethereum’s roadmap.

However, proof generation can be computationally expensive. If AI demand continues to consume advanced GPU supply, ZK infrastructure may face higher costs or longer deployment timelines.

This could accelerate three trends:

1. ASICs for proving
   Custom chips may become attractive for high-volume proof generation.

2. Protocol-level proof efficiency
   Teams may reduce circuit complexity, optimize state access, or use hybrid proof systems.

3. Prover marketplaces
   Networks may coordinate specialized operators that compete on reliability, speed, and cost.

The winning designs may be those that treat cryptography and hardware as a co-design problem, rather than separate disciplines.

Crypto Security in a Compute-Constrained World

A world of expensive compute also changes user security assumptions.

As AI agents become more common in crypto, users may delegate tasks such as portfolio monitoring, DeFi execution, cross-chain routing, tax classification, and governance participation. These systems may run on centralized servers, decentralized compute networks, or personal devices.

But the more automated the crypto experience becomes, the more important private key isolation becomes.

AI can help detect risk, summarize transactions, and automate workflows. It should not become an uncontrolled signer.

This is where hardware wallets remain relevant. A secure signing device separates decision-making from key custody. Even if an AI agent, browser extension, or cloud service is compromised, the private key should remain isolated from the internet-connected environment.

For users managing long-term assets, interacting with DeFi, or experimenting with AI-powered crypto tools, the key principle is simple: automation can assist, but signing authority should remain under user control.

What Builders Should Watch in 2025 and Beyond

The intersection of AI chips and blockchain infrastructure will likely become more important over the next few years. Builders and investors should watch several signals.

1. Cost of Proof Generation

If ZK proving costs decline, rollups and privacy-preserving applications become easier to scale. If hardware shortages keep costs high, adoption may concentrate among better-funded teams.

2. Growth of Prover Networks

A mature prover market could become as important to rollups as validators are to base-layer chains.

3. AI Inference on Crypto Rails

On-chain AI agents will need secure execution, reliable inference, and controlled signing. The wallet layer may become the user’s policy engine.

4. Hardware Supply Chain Diversification

The industry will pay more attention to chip design, packaging, memory bandwidth, and geographic concentration. Compute sovereignty will become a serious topic for both AI and crypto.

5. Efficiency-First Protocol Design

Projects that reduce computational waste may gain an advantage over those that assume unlimited hardware access.

The Bigger Picture: From Brute Force to Co-Design

Liao’s warning about a possible “avalanche” should not be read as a prediction that AI progress will stop. It is better understood as a warning that the old scaling playbook is becoming less reliable.

For crypto, the message is especially important. The industry is entering a phase where blockchains, AI systems, cryptographic proofs, and hardware infrastructure are becoming deeply connected. Future winners may not be the projects with the loudest narratives, but those that design across layers: protocol, cryptography, compute, security, and user experience.

The next frontier is not simply bigger chips or bigger models. It is better coordination.

And for individual users, that coordination should include secure self-custody. As crypto applications become more automated and AI-driven, protecting private keys becomes even more important. OneKey hardware wallets are designed to keep signing isolated from online environments, helping users maintain control while interacting with an increasingly complex on-chain world.