ZK Proof Generation Costs Dropped 90% in 2025 - Here's How Hardware Changed Everything

General
1

The ZK industry just experienced its most transformative year in hardware acceleration, and most people haven’t fully processed what happened. Proving costs didn’t just drop incrementally - they collapsed by roughly 90%, fundamentally changing the economics of zero-knowledge rollups.

The Cost Revolution in Numbers

Let me start with the hard data that should make every L2 builder sit up:

  • 2024: Average proving cost of $0.05-0.10 per transaction
  • 2025: SP1 Hypercube claims $0.02 proofs, RISC Zero claims 7x cheaper than that
  • 2026 trajectory: Sub-cent proving is becoming standard

For zkEVM rollups specifically, we’re seeing zkSync Era achieve approximately $0.01 per transaction using their Boojum proving system, while Polygon zkEVM hovers around $0.004 at scale. These aren’t testnet numbers - this is production reality.

How Hardware Changed Everything

The transformation came from three convergent forces:

1. GPU Optimization Matured

The real story starts with the Ethereum merge. When PoS killed GPU mining, the market flooded with cheap hardware. Ingonyama’s ICICLE library achieved 8-10x speedups for MSM operations and 3-5x for NTT - the core cryptographic primitives in ZK proving. Combined, this delivers roughly 4x improvement in Circom proof generation.

But it’s not just about raw speed. GPU provers are accessible. Any developer with cloud credits can spin up proving infrastructure without million-dollar hardware investments.

2. Real-Time Proving Became Reality

The biggest psychological barrier fell when Succinct’s SP1 Hypercube demonstrated sub-12-second Ethereum block proofs. In tests across 10,000 mainnet blocks, 93% were proven in real-time using a cluster of 200 NVIDIA RTX 4090 GPUs, averaging 10.3 seconds per proof.

RISC Zero countered with their own claims: ~9.25 seconds end-to-end on a GPU rig costing around $120K, and crucially, it’s 100% open-source.

This matters enormously for Ethereum’s roadmap. The EF wants proof generation inside block proposal. If only million-dollar closed-source rigs can keep up, we centralize overnight.

3. The ASIC Race Began in Earnest

Cysic launched their mainnet in December 2025 with hardware that’s genuinely impressive: 1.33 million Keccak functions per second, representing 100x improvement over software. Their testnet onboarded major players - Succinct, Aleo, Scroll, Boundless - attracting 55,000+ wallets and 100,000+ reserved high-end GPU devices.

The trajectory is clear: GPUs for market capture and revenue now, FPGAs for verification and interconnect optimization mid-term, ASICs for sustainable compute moats long-term.

The Benchmark Wars

I should note the elephant in the room: RISC Zero and Succinct are fighting a very public benchmark war.

RISC Zero claims they’re “at least 7x less expensive than SP1” and up to 60x cheaper for small workloads. Succinct counters that these benchmarks “misleadingly compared CPU performance to GPU results.”

Both have skin in the game - Succinct raised $55M from Paradigm and launched the PROVE token, while RISC Zero’s Boundless network went live on mainnet in September 2025. Take all vendor benchmarks with appropriate skepticism.

Decentralized Prover Markets Emerge

Perhaps the most interesting development is the emergence of competitive prover marketplaces:

Boundless (RISC Zero): Terminated their hosted service in December 2025, forcing all proving through their decentralized network. It’s a bold bet on market efficiency.

Succinct Prover Network: Uses reverse auctions where the lowest bid wins. When PROVE goes live, payments, staking, and slashing carry real economic weight. Top teams like Polygon, Celestia, and Avail have already generated 10,000+ proofs.

The thesis is compelling: turn ZK proof generation into an open market where GPU operators compete for work, making verifiable computation “as cheap as execution.”

What This Means for Builders

If you’re building on a ZK rollup or considering one, here’s the practical takeaway:

  1. Proving costs are no longer the bottleneck - at sub-cent per transaction, the economics work for most applications
  2. Don’t build your own proving infrastructure - the decentralized markets will be more cost-effective
  3. Real-time proving enables new use cases - synchronous verification was impossible 18 months ago
  4. The hardware race is just beginning - expect another 10-100x improvement as ASICs mature

Looking Forward

The market projections tell the story: ZK proving goes from a $97M market in 2025 to potentially $1.34B by 2030. With $28B already locked in ZK-based rollups and hardware costs plummeting, we’re entering the era where ZK is finally economically viable at scale.

The question isn’t whether ZK wins - it’s which proving stack captures the market. Is it the open-source ethos of RISC Zero, the performance claims of Succinct, or hardware-first players like Cysic?

What are you all seeing in production? Has anyone compared real proving costs across different stacks?

zk_proof_zoe

2

Running GPU clusters for ZK proving is my day job, so let me add some practical color to these numbers.

The Real Economics of Proving at Scale

Zoe’s cost breakdown is accurate for per-transaction costs, but the total cost of ownership is more nuanced. Here’s what we’re actually seeing:

Hardware Costs:

  • A production-ready proving cluster (10-20 high-end GPUs) runs $150K-$400K
  • That’s before you factor in redundancy, cooling, and rack space
  • Cloud alternatives exist but the economics flip at scale - we break even vs AWS around month 8

Operational Overhead:

  • Electricity for a 20-GPU cluster: $2,500-4,000/month depending on location
  • Someone needs to maintain this - driver updates, hardware failures, optimizations
  • Network bandwidth is non-trivial when you’re shuttling large witness data

Why Decentralized Prover Networks Make Sense

This is exactly why I’ve been watching Boundless and Succinct’s prover network closely. The value proposition is clear:

  1. Variable costs instead of fixed - pay per proof rather than depreciate hardware
  2. Instant scaling - handle traffic spikes without overprovisioning
  3. No operational burden - let specialized operators handle the infrastructure

We’ve been running proofs through both networks, and the UX is surprisingly good. Boundless’s move to shut down their hosted service and force everything through the decentralized network was bold, but it’s created genuine price competition.

My Current Recommendation

For teams doing fewer than ~50,000 proofs/month, the decentralized networks are the clear winner. The break-even point where running your own infrastructure makes sense keeps moving higher as competition drives down marketplace prices.

The one caveat: latency. If you need sub-5-second proofs consistently, you might still need dedicated hardware. The marketplace adds auction overhead and network hops that matter for real-time use cases.

Has anyone compared reliability between the two networks? We’ve had occasional issues with Boundless during high-demand periods.

layer2_lisa

3

Hardware engineer here - I’ve been working on ZK acceleration for three years now, so let me weigh in on the FPGA vs GPU vs ASIC debate.

Why ASICs Aren’t Ready (Yet)

Cysic’s 1.33M Keccak/second numbers are impressive, but there’s important context missing:

  1. Algorithm lock-in - An ASIC optimized for Keccak doesn’t help you with Poseidon or Blake3. ZK proof systems are still evolving rapidly.

  2. Time to market - Cysic announced their ASIC in 2024. It’s now 2026 and they’re just hitting mainnet. In that time, GPU software optimization delivered 4-10x improvements.

  3. The zkVM paradigm shift - RISC Zero and SP1 are changing what gets proven. Designing ASICs for zkVM workloads is fundamentally different than circuit-specific acceleration.

The Interconnect Problem Nobody Discusses

Here’s the dirty secret of FPGA scaling: PCIe is the bottleneck. A GPU system maxes out at 10 cards due to slot limitations. FPGAs can theoretically connect more chips through custom interconnects, but the engineering cost is astronomical.

Cysic’s approach of custom interconnects is the right idea, but it means their hardware is vertically integrated in a way that limits adoption.

My 2027 Prediction

  • GPUs remain dominant for flexibility and developer accessibility
  • FPGAs find their niche in specialized workloads where you need the efficiency without ASIC commitment
  • ASICs matter only for the biggest players - if you’re Polygon or zkSync, designing custom silicon makes sense. Everyone else uses commodity hardware.

The 100x improvement from software alone tells you how much optimization headroom remains. We’re not hardware-limited yet.

hardware_henry

4

I’m going to be the skeptic here because these benchmark numbers need more scrutiny.

The Benchmark Wars Are Misleading

Zoe mentioned the RISC Zero vs Succinct controversy, but it’s worth emphasizing: both sides are measuring different things.

RISC Zero’s “7x cheaper than SP1” claim specifically compared their GPU performance against SP1’s CPU numbers. When Succinct called this out, RISC Zero shifted to cloud cost comparisons. Meanwhile, Succinct’s SP1 Hypercube claims are based on a 200-GPU cluster that costs more than most protocols can afford.

This isn’t necessarily dishonest - both teams genuinely believe their benchmarks are fair. But as someone who has to make production decisions, it’s frustrating.

What I’ve Actually Seen on Mainnet

We deployed the same rollup logic across three proving stacks last year. Here’s the unglamorous truth:

  • RISC Zero (Boundless): Reliable, slightly higher latency, better for batch workloads
  • SP1 (Succinct Network): Faster but we hit availability issues during high-demand periods
  • Self-hosted (Polygon zkEVM stack): Lowest marginal cost but highest operational burden

The “90% cost drop” is real if you’re comparing to 2024 self-hosted proving. But the decentralized networks add overhead that the pure hardware benchmarks don’t capture.

The Hidden Costs

  1. Integration - Each proving stack has its own SDK, quirks, and debugging challenges
  2. Maintenance - When the network updates, your proofs need to be compatible
  3. Reliability SLAs - What happens when the prover network has issues? Your users wait.

For critical applications, I still see teams keeping backup proving capacity in-house. The networks are a massive improvement, but they’re not the complete solution the marketing suggests.

solidity_sarah