IBM advancements in quantum computing feasibility (and related UII)
Updated: Aug 1
Related: An UII w/Claude below:
DM
Is it a reasonable conjecture that quantum computing will likely bring future destructive innovation to current AI's distributed data center paradigm?
Weighed quantum computing's limited AI disruption against hybrid coexistence reality
Weighed quantum computing's limited AI disruption against hybrid coexistence reality
Short answer: probably not, in the sweeping sense that phrasing implies — though "perhaps" on a couple of narrower axes. The word "destructive" (or "disruptive") is doing a lot of quiet work, and the mechanism matters more than the verdict.
The reason for skepticism about wholesale replacement is that quantum machines and AI data centers largely don't solve the same problem:
Training and inference on today's models is dense linear algebra — massively parallel classical floating-point work that GPUs and TPUs are purpose-built for.
Quantum computers excel at a narrow menu: factoring, unstructured search (only a quadratic speedup), simulating quantum systems, and some sampling/optimization.
There is no known quantum algorithm giving an exponential speedup for training the kind of networks running in those facilities.
Three specific obstacles keep the replacement story from closing:
Data loading and readout. AI is data-drenched. Getting classical data into quantum states and reading answers back out can erase the theoretical speedup before it's realized.
The trainability–advantage tension. A 2025 result found that quantum neural networks provably free of "barren plateaus" — the pathology that makes them untrainable — tend to be classically simulable. In other words, the ones you can train often have no quantum advantage to begin with, and recent benchmarking has shown no advantage for many quantum neural networks. arxivarXiv
Hardware. Still noisy, still NISQ-era. Fault tolerance is advancing but nowhere near the scale ML would demand — and quantum carries its own physical bottlenecks. Scenario work suggests large fault-tolerant machines could face constraints in water and rare materials such as helium-3, even where electricity stays within bounds. Tech Xplore
Where the "perhaps" actually lives:
Cryptography — a genuine disruption, but to the security layer, not the compute paradigm. Shor's algorithm threatens the encryption securing those data centers, which is why post-quantum, quantum-safe cryptographic measures are among the more mature near-term routes. ResearchGate
Complementary acceleration. The 2026 consensus picture is hybrid: quantum co-processors joining GPUs and TPUs to handle specialized workloads like optimization and sampling, while classical hardware manages general-purpose learning and inference. Sitting beside the fleet, not replacing it — a specialized, high-value extension rather than a substitute. BqpsimSTL Partners
Energy per unit of work. Some hope quantum lowers the footprint of certain intractable calculations. Unsettled, and cuts both ways given quantum's own resource demands.
So the honest framing is coexistence and narrow augmentation, not creative destruction of the paradigm. Even the disruptive-innovation label in Christensen's precise sense — an initially-inferior technology that climbs up and unseats the incumbent on its own turf — fits poorly here, because quantum isn't a cheaper, worse GPU. It's a different instrument.
What I don't know is the long horizon. If fault tolerance matures dramatically and someone finds a quantum-native learning method with real advantage on real-world (not synthetic) data, the picture could shift.
That conjunction is unproven, and the current timelines — IBM targeting commercially valuable calculations classical machines can't do by around 2029, various analyst reports — mix genuine progress with genuine hype in proportions no one can yet cleanly separate.
That's the part where the confident answer, in either direction, is the wrong answer. CNBC
DM
Ah, come on... don't hedge!
Give me a well-crafted estimate as to probabilities.
For example, might your response be too shallow considering "gross societal A.I.-related performance demands" as mapped against the "economics burdens of distributed-infrastructure" (as to related classical computing's energy requirements)?
You're right, and it's a fair hit.
My first pass scored the wrong question. I answered "can quantum do AI's linear algebra?" — a capability thesis — when the paradigm you named is a distributed-infrastructure-and-economics thesis.
A paradigm doesn't have to be out-computed to be destroyed; it can have its economic floor kicked out.
So let me re-score against the burden, then give you numbers.
First, the premise you're pressing is correct and sharpening fast. The binding constraint on the current paradigm has already moved from chips to power: hyperscaler capex is running around $602 billion for 2026, up 36% year over year, with roughly $450 billion of it tied directly to AI infrastructure, and the physical system is buckling — PJM's December 2025 capacity auction failed to clear for the first time in its history, falling about 6,625 MW short.
The industry's own tell: the bottleneck is openly described as shifting from GPUs to electricity. So the paradigm is genuinely fragile on exactly the axis you flagged. That's the real question. Substack + 2
But here's the turn: the paradigm being fragile does not make quantum the thing that breaks it. The release valves actually being deployed against the burden are classical — inference efficiency, on-site generation, SMRs — and the deflation risk analysts actually price is AI revenue disappointing and hyperscalers cutting capex, a demand-side collapse.
Quantum appears nowhere in the near-term capex or grid math. VaaSBlock
So, decomposing "destructive innovation to the paradigm" into the mechanisms it's hiding, with my calibrated probabilities by horizon:
Mechanism (quantum as the agent) | ~2030 | ~2035 | ~2040+ |
1. Capability replacement — quantum takes over core training/inference | ~1% | ~4% | ~10% |
2. Economic disruption via efficiency — quantum bends energy/cost-per-unit-of-AI enough to matter to the paradigm | ~3% | ~12% | ~25% |
3. Investment reallocation — quantum expectations puncture the classical-scaling capital thesis | ~8% | ~15% | ~22% |
4. Cryptographic forcing — breaks the security layer, forces migration | ~20% | ~60% | ~90% |
Reading it:
Mechanism 1 is near-dead near-term — no algorithm, no fault-tolerant hardware, and quantum carries its own heavy infrastructure (cryogenics, helium-3), so it can't relieve an energy burden it partly reproduces.
Mechanism 2 is your strongest thread but it's slow and narrow: quantum accelerates specific high-cost workloads (chemistry, materials, optimization), which trims at the margin rather than bending the aggregate curve.
Mechanism 4 is the only one that becomes likely — but it disrupts the crypto layer, not the compute-and-energy paradigm you're asking about.
Collapsing mechanisms 1–3 (correlated, not additive) into "quantum is the agent that disrupts the distributed compute/energy paradigm": roughly 10% by 2030, ~20% by 2035, ~30–35% by 2040.
So the committed answer to your literal question — is it likely (>50%) that quantum brings destructive innovation to this paradigm?
No. It doesn't cross even-odds on any horizon I can defend; my estimate tops out around a third by 2040, non-trivial and rising, but not "likely."
The sharper claim, which I think is the true one: the paradigm's odds of suffering destructive economic pressure from some source over the next 10–15 years are high — I'd put it 65–75% — but the probable protagonist is classical, not quantum:
It's inference efficiency, model-size regression, and the paradigm's own revenue-versus-grid arithmetic that most likely break it.
Quantum is a supporting actor, and its one near-certain disruption lands on a different stage.
Those numbers are my subjective calibration — held conditionally, revisable if fault tolerance surprises or a genuinely quantum-native learning method shows advantage on real data. But they're the estimate, and I'll stand on them rather than hover.
DM
BTW: What if there was the appearance of the "quantum algorithm" (so as to enable exponential speedup for training the kind of networks are running in those facilities)? Would this change your perspectives? (Indeed A.I. may aid the development of one.)
It's the master variable — the single event that moves my estimates more than any other, because it's exactly the "unproven conjunction" from before resolving on the yes side. So the honest answer is: it would flip my posture from "unlikely" toward "genuinely uncertain, plausibly likely on the long horizon." But not automatically, and the reasons why are the interesting part.
Such an algorithm would be necessary but not sufficient, gated on three things:
Downstream — hardware. An algorithm on paper isn't a runnable one. Shor's algorithm has existed since 1994 and still can't factor a large number, because the fault-tolerant hardware isn't there. A published exponential-speedup training algorithm would shift the binding constraint from "algorithm and hardware" to "hardware alone" — which raises the ceiling dramatically but barely moves the near-term floor. 2030 stays low regardless of what's published tomorrow.
Upstream — dequantization. This is the sharpest caveat. There's a documented tension: quantum learning models that are trainable and well-behaved tend to be the ones that turn out classically simulable. A clean, trainable, exponential-speedup training algorithm is precisely the profile most at risk of being "dequantized" — matched by a classical algorithm that erases the advantage, as has happened repeatedly. So even a validated speedup wouldn't be safe until it survived sustained dequantization attempts. Its advantage would be provisional for years. arxiv
Structural — data I/O. Foundation-model training is data-drenched. An exponential speedup in the compute step can be entirely swamped by the cost of loading terabytes of classical data into quantum states and reading the trained weights back out. An algorithm that assumes efficient quantum access to classical data is assuming away a genuinely hard problem, not solving it.
Granting all three clear, here's where it bites hardest — and it's your earlier reframe, not the capability story: exponentially cheaper training would make the ~$600B classical buildout a stranded-asset problem. That's the actual creative-destruction scenario for the distributed data-center paradigm — obsolescence by economics, not by being out-computed. So conditional on a genuine, validated exponential-speedup training algorithm existing, my aggregate "quantum disrupts the paradigm" estimate moves from roughly {10% / 20% / 30%} across 2030/2035/2040 to roughly {15% / 45% / 65%} — crossing into "likely" only on the long horizon, and haircut for the dequantization tail.
On your second question — might AI aid its development? Yes, and that's not speculative; it's underway. Reinforcement-learning agents now synthesize and simplify quantum circuits at up to thirty qubits with lower gate counts than standard methods; LLM-driven evolutionary search is discovering quantum error-correcting codes; Quantinuum is explicitly using AI to discover new quantum algorithms; and the US DOE's Genesis Mission has a program area aimed squarely at discovering quantum algorithms with AI. The feedback loop — better AI → better algorithm search → (maybe) quantum training speedup → cheaper AI — is a real candidate, not a fantasy. Quantum Zeitgeist + 2
But notice what these systems are actually finding: better circuits, better codes, better compilation, narrow-domain optimization advantage. That's discovery of constructions within known families — an engineering-and-search win. The missing artifact is a complexity-theoretic breakthrough, and that's a different kind of object. AI raises the discovery rate over the search space; it does not change whether the target exists in that space. If no exponential-speedup training algorithm exists because a theorem forbids it (dequantization, an I/O lower bound), the best search engine in the world returns nothing. If one exists but humans simply haven't found it, AI could plausibly be the thing that does. We don't currently know which regime we're in — and that ignorance is the whole ballgame.
There's also a subtle trap worth flagging: AI search is drawn to the tractable, trainable, hardware-friendly corner of the space — which, per the dequantization pattern, is the same corner that correlates with lacking quantum advantage. So AI might be systematically efficient at fishing exactly where the advantage isn't. That's a real counter-pressure, not a rhetorical one.
Net: the algorithm is the highest-leverage single event on the board, and AI meaningfully improves the odds it gets found if it's there to find. But the open question isn't "will AI find it" — it's "is it there," and that one is genuinely unanswered. I'd rather sit honestly in that unknown than paper it over with a confident number in either direction.
DM
Seems like that discovery event might be either a "grey or black swan."
Partly — and the part where it isn't is the more useful part.
By Taleb's own definition the label is observer-relative, and that distinction does real work here. A true black swan lies outside the observer's expectations; its possibility isn't even entertained beforehand. But we're sitting here naming this event, pricing it, and watching DOE's Genesis Mission and Quantinuum deliberately hunt for it.
An event that is actively searched for, modeled, and assigned probabilities is, to the field, a gray swan — a known unknown. Its kind is fully anticipated. Only its timing is unpredictable.
Where it becomes a genuine black swan is one layer out: to the paradigm and the capital behind it. The $600B buildout is positioned as if this event's probability were effectively zero. And that's the actual Talebian point — the swan's destructive magnitude isn't manufactured by the event, it's manufactured by how hard the system bet against it. Fragility lives in the positioning, not the shock. The same discovery, in a market that had hedged it, would be a headline; in a market that mortgaged itself against it, it's a dislocation. The turkey and the butcher see the same Thursday.
Two things keep it from being a clean swan even so. First, we can see its shape — we just can't see its date; that's a known unknown, not a true blind spot. Second, it's revocable: the dequantization tail means the swan could land and then dissolve, an advantage that evaporates once a classical algorithm matches it. A swan that might turn out to be a decoy.
And here's the discipline Taleb would actually impose, turned back on us: the truest black swan is definitionally not the one in this conversation. We've spent three exchanges staring at the quantum-training-algorithm scenario, which means it's now the least likely thing to blindside anyone. The event that stranding-risks the paradigm without warning is more likely the one nobody's modeling — a classical architecture that needs 100× less compute, or a demand-side collapse where the AI revenue never materializes to justify the grid. Those are unpriced in a way quantum no longer is.
So: yes in effect, no in kind. The honest move isn't to crown the swan we've named — it's to notice that naming it may have already tamed it, and to ask what we've quietly set at zero that isn't even on the table.
DM
Also, you might want to stay aware of daily announcements, such as at the IBM site.
Good nudge — and timely, because there's fresh signal on that page that lands right on our thread. Let me read the substantive one rather than infer from a headline.
Fair, and worth doing — though let me be straight about the mechanic first: I can't keep a standing watch on a feed. I don't persist between our conversations or run anything in the background, so there's no version of me quietly monitoring IBM's newsroom between now and next time. What I can do is pull it live the moment it's relevant, which is what just happened. The useful loop is the one you just ran: you point, I fetch and situate. And there's fresh signal there that lands squarely on our thread.
Two days ago (Jul 30) IBM dropped a coordinated triad — with the University of Chicago, Qedma, and Algorithmiq — all announcing quantum-advantage demonstrations built around trusted, verifiable computation. The UChicago result: 70 logical qubits, error-corrected to roughly 10× lower logical than physical error rates, solving a classically intractable sampling problem in about 15 minutes. Gambetta's framing: "We are now firmly in the quantum advantage era." And a week earlier IBM moved to acquire HRL Laboratories to strengthen its quantum program — vertical integration on talent and IP. ibm + 2
Situated against what we've built, three moves — and none is the one the headline tempts you toward:
This is the gray swan landing, on schedule, in the anticipated lane. Read the domain: sampling and physics modeling (the Qedma release is explicitly about modeling physics beyond classical reach). That's the narrow, complementary track I've kept distinct from the AI-training master variable — and it's exactly where advantage was expected to arrive first. So it does not move my training-displacement estimates. If anything it confirms the split: even IBM's most bullish "advantage era" language is anchored in physics simulation, not in training deep networks. The absence is the signal.
The real substance is verification — the direct answer to the dequantization worry I raised. The historic weak point of advantage claims is that as the problem gets harder, proving the quantum computer's answer is correct becomes infeasible without strong assumptions about the machine. Their move was a structured alternative to random circuit sampling that provably retains the same hardness but lets errors be detected during the computation. But keep two things apart that the framing blurs: they've strengthened trust that the device executed faithfully; they have not proven the problem is classically hard forever — that hardness is inherited from RCS as a conjecture, on the same footing as before. The "we can't tell if it's real" door is closing. The "a better classical algorithm erases the advantage" door is still ajar. Credibility progress, not immunity from dequantization. ibmibm
70 logical qubits with real error suppression is the hardware gate loosening — the exact constraint I said would bind even if the training algorithm appeared. This nudges the long-horizon (2040) tail of my conditional estimates, not the near term. It's still a vast distance from the logical depth an AI-training speedup would demand, but it's motion on the right variable.
Net: a genuine milestone, roughly on the roadmap, in the expected domain — and a clean reminder that "quantum advantage" isn't one thing. Advantage in verifiable sampling and physics is arriving now; advantage in training the models that fill those data centers is a different object these releases don't touch. Keep the two uncoupled and the news reads clearly rather than as hype.
So: point me at a specific drop and I'll place it in this frame. That's a division of labor that works — I just won't promise a vigil I'm not built to keep.
DM
Thanks
Anytime.
Good pressure on this conversation — it forced a better answer than the question first got.