The fence at 87th Street

Originally published on billfaruki.substack.com on July 25, 2026. Republished on MindHYVE.ai with light canon edits. The original post's photograph of the site is not reproduced here; see the original for the image.

My daughter sent me a photograph of a construction sign.

Red and white, bolted to a gray timber frame on a patch of grass off 87th Street. Hard hats, safety glasses, gloves, high-visibility vest. A phone number for a man named Bob. Emergency Access Gate A, 3733 E. 87th St.

She had gone to see what used to be a steel mill. She wanted to know what was going on.

Behind that fence is the former U.S. Steel South Works — 440 acres of Chicago lakefront that sat empty for three decades after the mills went cold. It is now the Illinois Quantum and Microelectronics Park, anchored by PsiQuantum, with IBM, Infleqtion, Diraq, and Quantum Machines committed to the campus. Ground broke in September 2025. This phase finishes in 2027. A 65,000-square-foot building meant to house what may become the first utility-scale fault-tolerant quantum computer in the United States is already most of the way up.

So I started typing an answer.

And the first sentence I wrote was wrong.

It was a version of a sentence you have read a hundred times:

Quantum computing is right around the corner, and it will unlock breakthroughs in AI.

Comfortable.

Confident.

Backwards.

Neural networks are dense linear algebra over classical data. That is precisely the workload GPUs were built for and precisely the workload quantum hardware is bad at.

Quantum machine learning has two structural problems that have resisted a decade of serious effort.

The first is data loading: getting classical data into a quantum state often costs more than any speedup you would recover from operating on it.

The second is barren plateaus, where the gradients in variational circuits flatten toward zero as the circuit scales, and training simply stops working.

There is no demonstrated quantum advantage on a real machine-learning task. Not a contested one. Not a small one.

Meanwhile, the arrow runs hard in the other direction.

The most consequential recent work in quantum error correction uses machine learning to decode surface codes — AlphaQubit, from one of the major AI labs, being the visible example. AI is calibrating qubits, shaping control pulses, reading out error syndromes.

Today, AI is a quantum technology far more than quantum is an AI technology.

Which means the question my daughter actually asked is better than the one I was about to answer.

Not what will this do for our models?

What will it do that our models can then learn from?

Strip away the marketing and there are three tiers.

Tier one is real and structural.

Simulating quantum systems.

Molecular ground states, catalytic transition states, correlated-electron materials, nitrogen fixation, high-temperature superconductivity.

This was Feynman's argument in 1981 and it remains the only application class where the advantage comes from the physics rather than from hope.

Nature is quantum; simulating it on classical hardware costs exponentially; simulating it on quantum hardware does not.

Tier two is real but conditional.

Cryptanalysis.

Shor's algorithm works — the only question has ever been resource cost, and that cost is collapsing faster than the hardware is improving.

In 2019, Craig Gidney and Martin Ekerå estimated that factoring a 2048-bit RSA key would take twenty million noisy qubits and eight hours.

In 2025, Gidney revised his own number: fewer than one million noisy qubits, in under a week.

A twentyfold reduction in the hardware requirement, achieved without building anything — purely through better arithmetic and better error correction.

Sit with that.

The bar is falling toward the machine at the same time the machine is climbing toward the bar.

Almost every timeline argument you read accounts for only one of those motions.

Tier three is mostly narrative.

Generic optimization — logistics, portfolios, scheduling, the slide in every quantum pitch deck.

Grover's algorithm delivers a quadratic speedup, and quadratic speedups are routinely consumed whole by error-correction overhead.

Leave this tier out of any serious conversation.

Your credibility on tiers one and two is worth more than the extra market you buy by claiming tier three.

Here is the thesis I would put my name to.

Quantum computers will not train better models.

They will produce better ground truth.

There are chemical and material systems we cannot compute at any price.

Not slowly — not at all.

The strongly correlated electron problems underneath catalysis, room-temperature superconductivity, and industrial nitrogen fixation are exactly the problems classical methods approximate badly and quantum methods address natively.

When those calculations become routine, they generate training data that has never existed.

And then classical AI does what classical AI is extraordinary at: learning the surface of an expensive function so it can be evaluated cheaply a billion times.

Run the quantum simulation on a few thousand hard cases.

Train a classical model on the results.

Search a chemical space of 10⁶⁰ candidates in an afternoon.

Send the twelve interesting ones back to the quantum machine for verification.

The breakthrough is not a quantum language model.

It is a classical model trained on chemistry no one has ever been able to compute.

The quantum computer is an instrument — closer to a synchrotron or a cryo-EM than to a faster GPU.

We do not ask what the telescope does for the astronomer's arithmetic.

We ask what it lets her see.

That reframe changes what to watch for.

Not qubit counts.

Not benchmark stunts on contrived sampling problems.

Watch for the first published molecular calculation a pharmaceutical company acts on commercially.

That is the starting gun.

Now the part almost nobody writes, and the part your CFO should care about this fiscal year.

The mechanism is harvest-now-decrypt-later.

An adversary captures encrypted traffic today and stores it against a machine that will exist in a decade.

If the material has a long secrecy half-life — health records, financial positions, source code, negotiation history, state secrets, anything under a legal retention obligation — then the encryption protecting it in 2026 is already inadequate.

The exposure was created the moment the packet crossed the wire.

Which is why NIST finalized the post-quantum cryptographic standards in August 2024 and why federal migration timelines are now fixed in policy.

Every CISO in a regulated industry owns a multi-year cryptographic inventory and migration with zero attached revenue.

That is not a future cost.

It is in this year's budget, being spent, against a threat that does not yet exist.

The economic impact of quantum computing arrives years before the computing does.

That asymmetry — a defensive clock running well ahead of the offensive one — is the most underpriced fact in the entire field.

Suppose the machine behind that fence works.

It hands you a molecular ground-state energy.

How do you know it is right?

The problems where quantum wins are, by construction, the problems classical computers cannot check.

You cannot validate the answer by recomputing it the old way — the whole premise is that the old way is intractable.

So the field will have to build a trust architecture from scratch:

• Cross-platform replication on independent hardware.
• Wet-lab confirmation loops.
• Statistical certification protocols.
• Provenance records for computational results.

This is not a technicality.

It is a genuinely new epistemic problem, and it will produce new institutions, new standards bodies, and a professional category that does not yet have a name.

Anyone who has spent time building reasoning systems recognizes the shape of it immediately:

The hard part was never generating the answer.

The hard part is generating the answer together with the evidence that lets a qualified human accept or reject it.

Quantum computing is about to meet that problem at industrial scale, and it will meet it where the rest of us do — at the boundary where a machine's output becomes a person's decision.

Something changed in how this gets funded, and it is a better maturity signal than any technical milestone.

The Chicago campus is backed by roughly $500 million in Illinois state commitment, with the anchor tenant's build financed by Blue Owl Capital funds.

Clayco is general contractor.

Delivery is 2027 for phase one and 2028 for the campus.

Advocate Health is putting a 53-bed hospital on the site.

That is project finance.

Infrastructure and real estate underwriting, with milestone draws and completion risk and a defined cost of capital.

It is not venture capital, which is what a field looks like when the question is still does the science work.

When a technology moves from equity risk to construction risk, the market has quietly answered the first question and gone on to the second.

You can read that off the balance sheet earlier than you can read it off the physics.

Dilution refrigerators, a dedicated cryogenic plant, industrial-scale power, water, and rail-and-lake logistics.

That is why this campus is on a decommissioned steel mill and not in a suburban office park.

Quantum computing has the site requirements of a foundry, and it is going to the places that were built for foundries.

Which brings me back to the fence, and to why I think a kid standing in front of it asked the right question in the wrong century.

South Works employed this neighborhood by the thousands.

Whole blocks were built on the assumption that the mill would carry families for generations.

Then it closed, and the land sat vacant for thirty years, and the people who stayed watched a field of weeds where their parents' work used to be.

That is the memory this campus is being built on top of.

The new industry has a thin, credentialed, extremely well-paid labor market concentrated in a handful of cities — Chicago, Boulder, Delft, Sydney.

Local organizers have pushed for a ballot referendum on the project.

Their questions are about environmental impact and about who actually gets the jobs.

Those are not obstructionist questions.

They are the same questions the last generation on this ground should have been allowed to ask before the answer was decided for them.

Whether quantum keeps a better version of the promise steel made is a governance problem, not a physics problem.

Governance is the thing this industry is least practiced at, and the thing that will determine whether the neighborhood remembers 2028 the way it remembers 1992.

Three things, if you run a company and are trying to decide what to do about any of this.

Start the cryptographic migration now.

It is the only item with a hard, already-running clock.

Inventory what you encrypt, how long it has to stay secret, and what algorithms protect it.

Unglamorous, expensive, non-optional.

If you are in chemistry, materials, pharma, or energy, build the data pipeline.

The bottleneck when quantum simulation becomes routine will not be access to the machine.

It will be whether you can turn expensive calculations into trained models and trained models into decisions.

That infrastructure is classical, buildable today, and work you should be doing anyway.

Ignore qubit counts.

They are the field's vanity metric.

Watch logical error rates, resource estimates for real problems, and the first commercial decision made on a quantum-computed result.

And be careful with right around the corner.

The Chicago facility targets 2028.

Utility-scale fault tolerance is a goal with a date attached, not a schedule with a guarantee.

Every serious person in this field will tell you that.

Every unserious one will not.

Something is being built on the ground where the mill was.

It is a real machine, not a metaphor, and it will do a small number of things no machine has done — mostly having to do with understanding how molecules hold themselves together, which turns out to be the bottleneck under medicine and energy and materials both.

It will not think.

It will not make our thinking machines think better.

What it will do is see a part of the physical world we have never been able to look at directly, and hand what it sees to people — and to the ordinary, unglamorous, classical software that has to turn seeing into deciding.

That building will be finished around the time she is choosing what to study.

Someone will have to decide what those first calculations get pointed at, and whether the answers can be trusted, and who in that neighborhood gets to work on them.

Those are not questions the machine answers.

Those are the questions left over after it does.

She asked what was going on.

That is the honest version.

First published on Bill Faruki's Substack on July 25, 2026. Lightly edited for house style. Read the original.