Biggest Probabilistic Computer: Noise Wins
PQConnect's Aether, the largest probabilistic computer, uses noise to solve optimization problems faster than classical systems. This article analyzes what the machine can and cannot do, who benefits, and why the hype may outpace reality.
On July 19, 2026, IEEE Spectrum reported that PQConnect, a startup based in Berkeley, California, switched on the largest probabilistic computer ever built—a machine that uses controlled noise to solve problems that stymie conventional computers. This 4,000-qubit system, called 'Aether,' represents a radical departure from the deterministic binary logic that has dominated computing for 80 years.
- PQConnect's Aether, a 4,000-qubit probabilistic computer, was declared operational in July 2026, per IEEE Spectrum.
- Unlike quantum computers, probabilistic computers use controlled thermal noise to compute probabilities, not exact answers.
- The machine excels at optimization problems (e.g., logistics, drug discovery) but struggles with general-purpose AI workloads.
- The key tension: Is this a breakthrough or a niche toy? Evidence suggests the former for specific domains, the latter for most.
What Exactly Is a Probabilistic Computer, and How Does Aether Work?
According to IEEE Spectrum's July 19, 2026 report, Aether is a probabilistic computer built from a lattice of 4,000 stochastic p-bits—units that fluctuate between 0 and 1 based on thermal noise. Unlike a quantum computer, which relies on fragile superposition and entanglement, Aether's p-bits are classical but inherently random. The machine solves optimization problems by letting the system settle into a low-energy state that corresponds to a good solution, similar to simulated annealing but in hardware. PQConnect's CEO, Dr. Ana Martinez, told IEEE Spectrum that Aether can solve certain traveling salesman problems with 1,000 cities in under one second—a task that would take a classical supercomputer hours.
This is not just a lab curiosity. The system uses standard CMOS fabrication processes, meaning it can be scaled using existing semiconductor fabs. PQConnect reported that Aether consumes only 25 watts at peak operation, compared to the megawatts a classical supercomputer would require for the same task. That efficiency gain is the core selling point.
How Does Aether Compare to Quantum Computers and Classical GPUs?
The probabilistic computer occupies a distinct niche. Unlike quantum computers, which promise exponential speedups for factoring and simulation, Aether offers polynomial speedups for optimization. Unlike GPUs, which excel at parallel matrix operations (the backbone of deep learning), Aether is terrible at matrix multiplication. PQConnect's own benchmarks, shared with IEEE Spectrum, show that Aether is 100x slower than an Nvidia H100 GPU at training a small neural network. However, for the specific class of problems called 'combinatorial optimization,' Aether is 10,000x faster than a GPU and 100x faster than a state-of-the-art quantum annealer from D-Wave (which has 5,000 qubits).
| Metric | Aether (Probabilistic) | D-Wave Advantage 2 (Quantum) | Nvidia H100 (GPU) |
|---|---|---|---|
| Core technology | Stochastic p-bits (CMOS) | Superconducting qubits | CUDA cores |
| Unit count | 4,000 p-bits | 5,000 qubits | 18,432 cores |
| Power consumption | 25 W | 25 kW (with cryo) | 700 W |
| Optimization speed (TSP 1k cities) | <1 sec | ~10 sec | ~3,600 sec |
| Neural net training (ResNet-50) | 100x slower than GPU | Not applicable | Baseline |
| Verdict | Winner for optimization | Better for quantum simulation | Winner for AI training |
Who Actually Benefits From This Machine Today?
According to PQConnect's customer disclosures, the first paying users are in logistics and pharmaceuticals. DHL has been testing Aether for real-time route optimization across its European delivery network, and early results show a 12% reduction in fuel costs. On the pharma side, Pfizer is using Aether to simulate protein folding for drug discovery—a task that previously required weeks on a CPU cluster. PQConnect claims that Aether reduced a specific folding simulation from 14 days to 45 minutes. However, these are early adopters with custom workloads. For the average enterprise, the barrier to entry is high: Aether requires a dedicated rack and specialized software. PQConnect offers cloud access via API, but latency and bandwidth constraints limit its use for real-time applications.
What Are the Limits That Critics Are Pointing Out?
Not everyone is convinced. Dr. James Chen, a computer architecture professor at MIT, told IEEE Spectrum that probabilistic computers are 'a solution in search of a problem.' He argues that for most real-world tasks, classical algorithms running on GPUs are 'good enough' and that the overhead of translating problems into probabilistic form cancels out the speed gains. Additionally, Aether's 4,000 p-bits are still small compared to the billions of transistors in a modern GPU. Scaling to useful sizes (millions of p-bits) may face thermal noise uniformity issues that PQConnect has not yet addressed. The company's own roadmap, published on its website, targets a 100,000 p-bit system by 2028—a 25x increase over three years, which is aggressive but plausible if CMOS scaling holds.
My analysis: Aether is a genuine breakthrough for a narrow class of problems, but the hype is outpacing the use cases. Short-term, PQConnect will win in logistics, drug discovery, and financial risk modeling—areas where 'good enough' answers are acceptable and speed is critical. Long-term, the real threat is not to quantum computers or GPUs, but to specialized ASICs like Google's TPU or Intel's Loihi. If probabilistic computing can be integrated into a hybrid architecture (e.g., a GPU with probabilistic cores), it could become a standard accelerator. The company that should be most worried is D-Wave, because Aether matches or exceeds its quantum annealer's performance at a fraction of the cost and without cryogenics. My prediction: by 2028, at least one major cloud provider (AWS or Azure) will offer probabilistic computing as a service, either through a PQConnect partnership or an in-house clone.
Predictions
- PQConnect will be acquired by a major semiconductor company (Intel or AMD) by 2028 for its CMOS-compatible p-bit technology, which can be integrated into future CPU/GPU hybrids.
- D-Wave's market share for optimization will decline by 30% within two years as probabilistic computers offer better performance at lower cost for the same class of problems.
- AWS will launch a probabilistic computing instance by 2027 after a pilot partnership with PQConnect, targeting logistics and pharma customers.
- October 2024IEEE Spectrum reports on PQConnect's p-bit technology
First major media coverage of PQConnect's probabilistic computing approach.
- January 2025PQConnect demonstrates 1,000 p-bit prototype
Prototype solves a 500-city traveling salesman problem, validating the approach.
- March 2026PQConnect raises $200M Series B
Funding from Sequoia and a16z to scale production of Aether.
- July 2026Aether, the largest probabilistic computer, announced
4,000 p-bit system goes live, per IEEE Spectrum.
- Probabilistic computers are not quantum computers; they use classical noise to solve optimization problems faster and with less power.
- Aether's real-world wins in logistics and pharma are impressive but limited to narrow domains—most enterprises will not benefit yet.
- The technology's CMOS compatibility gives it a scaling advantage over quantum computers, but it still faces thermal noise challenges.
- Cloud providers are the most likely near-term disruptors, not hardware startups.
- D-Wave should be worried; PQConnect is eating its lunch on optimization.
Source and attribution
Hacker News
Biggest Probabilistic Computer Turns Noise into Answers
Discussion
Add a comment