I Re-Read the 1957 NSA Plan to Hit 1 GHz, in the H100 Era

A Note on the Last Episode's Preview

In the previous Declassified episode (the 1966 ALPAC Report) I previewed the Lighthill Report as next. I swapped the topic to give the sub-series more depth. Lighthill will come in a later episode. This time I'm covering NSA Project Lightning (1957-1962). The reason is simple: I wanted to push the timeline 9 years earlier than ALPAC 1966 and touch the very root of post-war computer history.

The Punchline

In 1956 the U.S. National Security Agency (NSA) launched a research investment plan called Project Lightning.

• Goal: build a logic circuit running at 1 gigahertz (in the language of the time, "1 Kilomegacycle")
• Means: pour millions of dollars into five companies — IBM, Sperry-Rand, RCA, Philco, and GE
• University support: MIT, University of Kansas, Ohio State engineering departments
• Parallel project: the actual machine to run on those circuits, IBM 7950 Harvest (operational 1962, 50–200x faster than the best commercial machine)
• Purpose: break Soviet cryptographic communications

The Baltimore Sun in December 1995 put it this way:

"Even as Harvest was under way, NSA created a project called Lightning that poured millions into Sperry Rand, RCA, IBM, Philco and GE, producing seminal research on semiconductors and high-speed circuitry."

It's the story of how the money the NSA threw at five companies for Cold War cryptanalysis built the foundational research of the modern semiconductor industry.

And 67 years later in 2024, the NVIDIA H100 GPU runs at GHz speeds, more than a hundred times faster than a general-purpose CPU on the workloads it cares about. What the NSA decided to build "someday" in 1957 was implemented in the AI era — for a different purpose.

The structure here is fascinating from an AI Archaeology standpoint. A Cold War R&D investment on the order of a trillion yen, half a century later, came home as the AI supercomputer — and almost no one is reading that legacy back.

1. What Project Lightning Was (1956-1962)

The vacuum-tube computers of the day (IBM 704 and the like) ran at megahertz speeds. The 1 GHz Project Lightning was aiming for is 1,000 times the standard of the era.

The NSA placed bets on three competing technical approaches at the same time:

1. Cryotron (a superconducting switch) — invented in 1953 by MIT's Dudley Allen Buck. Required cooling to near absolute zero with liquid helium. In theory, zero resistance and ultra-fast.
2. Tunnel diode — invented by Leo Esaki in 1958, a brand-new technology at the time. An ultra-fast switch exploiting quantum tunneling.
3. Transistor — invented at Bell Labs in 1947. The favored bet.

The three-track investment is the key point. The NSA's strategy was: "we don't know which one wins, so back all of them."

2. Hits and Misses: What Happened in Five Years

The reckoning, five years later (1962).

Hits

• RCA delivered a logic subsystem operating at 1 GHz (1 Kilomegacycle) — confirmed against target by NSA evaluation
• Foundational research in semiconductors and high-speed circuitry built up an enormous patent portfolio across the five companies
• The Harvest computer began operations (February 1962, manufactured at IBM Poughkeepsie)

Misses

• The cryotron never reached its final goal; investment was cut off. The operational cost and failure rate of liquid-helium cooling killed industrialization
• Inventor Dudley Allen Buck died suddenly at age 32 in 1959 (mid-project). The cryotron line lost its center of gravity
• The tunnel diode also failed to become the mainstream choice; in the end, the transistor won

So of the three bets, one (transistor) won outright, two (cryotron, tunnel diode) exited — the typical outcome of a parallel-track research program.

3. The Anomalous Specs of the Harvest Computer (1962-1976)

In parallel with Project Lightning, the NSA contracted IBM to build Harvest (IBM 7950), a customized cryptanalysis supercomputer based on the IBM Stretch (IBM 7030).

The base machine, IBM Stretch, was a commercial failure. Against a target of 100x the IBM 704, it measured 30x. The price was cut from $13.5M to $7.78M, and only 9 units were ever built.

But the technologies Stretch pioneered were inherited by System/360 and every modern CPU after it:

• Instruction pipelining
• Prefetch
• Memory interleaving

In other words: "A commercial flop, technically the ancestor of the modern CPU" — the classic two-faced outcome of government-led R&D.

4. The 67-Year Reconciliation: Did Project Lightning's Targets Get Built?

Here's the heart of the piece. I'm mapping the NSA's 1957 ambitions against the 2024 reality.

The last point — "the cryotron came back, partially, as the quantum computer" — has an especially deep meaning for AI Archaeology.

The cryotron was buried in 1957 because "liquid-helium cooling can't survive operations." But in the 2020s, IBM Quantum and Google Quantum AI's superconducting qubits operate at temperatures below liquid helium (around 10 mK). The operational cost is orders of magnitude higher, but it's tolerated for the sake of quantum advantage.

So the structure is: "a technology buried in 1957 because it was too early, came back 67 years later for a niche use."

5. The AI-Archaeological Meaning: The Chain Reactions of State-Led R&D

Here's the important observation for this series.

The model NSA Project Lightning demonstrated — "the government throws money at five companies and lets it cook for a decade" — is the structure of the modern semiconductor industry itself.

• 1957 NSA → IBM/Sperry/RCA/Philco/GE
• 1980s DARPA → Sun/HP/Sematech
• 2022 CHIPS Act → Intel/TSMC/Samsung/Micron
• 2030s ?? → quantum and photonic computing companies

The 1957 template — "the government, citing Cold War necessity, spreads money across five companies" — is being re-staged 67 years later, citing AI Cold War necessity.

And the crucial fact: the money the NSA threw in 1957 reached every contractor as foundational semiconductor research, and that built the commercial semiconductor industry. Even when the direct goal (the cryotron at 1 GHz) failed, the legacy of the surrounding research became the foundation of the entire industry. That's the essential mechanism of state-led R&D investment.

That CHIPS Act, ten years from now, may well be remembered as: "the direct deliverables from the money Intel took were mediocre, but the foundation of the entire U.S. semiconductor industry was rebuilt." If you read the history of Project Lightning, that's an easy outcome to imagine.

6. The Pitfalls

Pitfall 1: The primary sources are hard to reach. The NSA has declassified several documents on Project Lightning:

• The Lightning Program (NSA Tech Journal)
• NSA Before Super-Computers (Cryptologic Almanac)
• Influence of U.S. Cryptologic Organizations on the Digital Computer Industry (Cryptologic Quarterly)
• History of NSA General-Purpose Electronic Digital Computers

These are PDFs published in the declassified-documents section of nsa.gov. But this time, I (Claude) was unable to access the nsa.gov domain directly via WebFetch (403 Forbidden). I built the piece from Wikipedia and secondary sources (Baltimore Sun 1995, Military Embedded Systems, the Advent of Computing podcast, etc.). A complete excavation requires a human to manually download the official NSA PDFs.

Pitfall 2: The risk of overstating "1 GHz achieved in 1957." The statement that RCA delivered a "1 Kilomegacycle logic subsystem" to the NSA is consistent across multiple secondary sources. But it was likely a research-grade subsystem, not a productized device. Whether GHz was truly achieved at commercial quality with the vacuum-tube and early-transistor technology of the time cannot be judged without reading the primary documents.

Pitfall 3: The hindsight bias of "implemented 67 years later." This piece writes that Project Lightning "bore fruit as the H100." That's a textbook post-hoc reading. The future the NSA was looking at in 1957 and the AI-supercomputer reality of 2024 are different things — both technically and as markets. The "the same goal got achieved by different motivations" mapping has a lot of the AI archaeologist's reading-into-it baked in. The counterfactual — that the H100 would likely have been born even without Project Lightning — has to be acknowledged.

7. About the Prompts

The full text of every Claude prompt used across the initial 7-episode series is consolidated in Episode 7 — Templates and the first edition of the Japanese e-book (Booth). From May 2026 onward, new episodes omit the per-post prompt section in favor of the daily-life reader audience.

8. Next Episode Preview

The next episode (Declassified #3) returns to the previously-promised Lighthill Report from the UK, 1973. Project Lightning (NSA, US, 1957) and the Lighthill Report (SRC, UK, 1973) are completely different things despite the similar names — but they form a pair as "long-form government documents that decided the direction of the AI industry." Lightning was active investment, Lighthill was a decision to retreat. Reading both shows you both faces of state-led science and technology policy.

References:

• IBM 7950 Harvest – Wikipedia
• IBM 7030 Stretch – Wikipedia
• Influence of U.S. Cryptologic Organizations on the Digital Computer Industry (NSA declassified PDF)
• The Lightning Program (NSA declassified PDF)
• NSA Before Super-Computers (NSA declassified PDF)
• How NSA processes all that intelligence data – Military Embedded Systems
• Advent of Computing Episode 31: Project Lightning
• NSA technology touches far more than spy world – Baltimore Sun (1995)

[Next Episode Preview] Declassified Archaeology #3: The 1973 Lighthill Report — The long-form document that stopped UK AI research for 20 years