The collapse of the Lisp machine market

Filed as moment, and moment is right — but the moment is a fiscal year, not a day, and the entry has to say so before it says anything else. The house definition, set by eliza-1966 when it argued its own way out of the kind and used again by clippy-1996, is a date on which something visibly happened in public: Dartmouth, Lighthill, Deep Blue, the LLaMA weights. There is no such date here. Nobody experienced 1987 as a day. What the record holds instead is a revenue line that turned over: Symbolics' consolidated product revenue was $101.6 million in fiscal 1986, $82.1 million in fiscal 1987, and $55.6 million in fiscal 1988, and the year label on this entry is the year the sign flipped. That is a real, checkable, public event with a date attached to it — the company's fiscal year ended 30 June — but it is an event of a different texture from a chess match, and a reading that cites this entry should know it is citing an accounting fact rather than a scene.

The proposal line is wrong in a checkable way, and correcting it is half the value of the entry. canon/proposals.md describes this as "a hardware and consulting industry built on one AI paradigm vanished in about a year." The hardware industry did not vanish in about a year. Symbolics lost money for six more years and filed for Chapter 11 in January 1993. The consulting and tools side lasted longer still: Teknowledge went public in 1986, sold itself to American Cimflex by 1989, was spun back out, became a DARPA contractor, was partly bought by Intuit, and did not file for bankruptcy until December 2013 — twenty-six years after the "collapse." Carnegie Group grew through the winter to more than 300 people and over $30 million in revenue, went public around 1996, and was sold in 1998. "Vanished in about a year" is the version of 1987 that gets told; it is not the version the filings support, and the gap between the two is precisely the thing a reading needs to be careful about when it reaches for this analogy.

Why not interpretation. The strongest alternative, and it nearly wins. What actually descends from 1987 into 2026 is arguably not the event but the frame — "AI winter," a phrase coined at AAAI in 1984 and now the standard lens through which every wobble in an AI market is read. An interpretation entry on the winter narrative would be a legitimate and useful thing. But it would be a different entry with a different id, and the id here names the collapse. The collapse is an event; events are moments. The frame appears below as something this entry warns about rather than something it is.

Why not limit. Nothing here is proved, and the reason to say so out loud is that "AI winters are periodic" is deployed constantly as though a law had been established. Two contractions in twenty years is not a cycle; it has no period, no mechanism that has been isolated, and no statement of the conditions under which it recurs. Filing this beside Gödel and Turing would lend it exactly the authority that godel-incompleteness-1931's section 4 exists to strip away.

Why not prediction. Also tempting, because there is a clean dated forecast at the centre of this — Minsky and Schank's 1984 warning, due 1987 — and it is graded in section 3 below. But the prediction is a claim about the event, not the event; if it deserves its own entry it should get one, and this one should not annex it.

descends_from holds one id, and the link is documented rather than inferred. MYCIN's rules were separable from its interpreter, and removing them left EMYCIN, the first expert-system shell. That is the standing claim in mycin-1976, which already names EMYCIN as "the direct ancestor of the commercial shells that drove the 1980s expert-systems industry, and with it the boom that ended in the second AI winter." What makes the edge a checked fact rather than a plausible story is Denny Brown, who ran the education and later the applications side at Teknowledge, saying it in the Computer History Museum's own recording: "We were in the MYCIN/E-MYCIN historical. E-MYCIN became what we called KS300, which was a redo and extension of E-MYCIN, which then was marketed as S.1." S.1 was Teknowledge's product. One of the four companies whose collapse this entry is about was selling a direct descendant of a Stanford dissertation, and said so on the record thirty-two years later.

Two ancestors I would want and cannot have: lighthill-1973, which is the first winter and is proposed but not written, and a DENDRAL entry, which is neither written nor proposed — mycin-1976 flags the same gap. If either is ever written, this entry should list it. Note what is not an ancestor: the Japanese Fifth Generation project, which is usually placed in this causal chain and is better understood as a parallel case — it also failed, at roughly $500 million over about a decade, but the American boom it helped provoke was funded and built independently of it.

What it is

Through the first half of the 1980s a commercial industry assembled itself around one AI technique. The technique was the expert system: domain knowledge written as IF/THEN production rules, kept separate from a general engine that applied them. The commercial proof was XCON (internally R1), written in OPS5 by John McDermott at Carnegie Mellon from 1978 to configure orders for Digital Equipment Corporation's VAX systems. XCON worked. The savings figure varies by source in a way worth noticing — Crevier reports it as $40 million over six years of operation, while other accounts give $25 million a year — but nobody disputes the direction. By 1985, corporations worldwide were spending over a billion dollars on AI, most of it on in-house AI departments.

An industry grew to supply them, in two layers.

The hardware layer was the Lisp machine: a workstation with a tagged architecture, hardware support for garbage collection, a large virtual address space and a famously good integrated development environment, designed from the silicon up to run Lisp. Symbolics was formed in 1980 by 21 founders, most out of the MIT AI Lab, and shipped its first product in 1981; Richard Greenblatt, who had built the original CADR at MIT, founded the rival Lisp Machines Inc.; Texas Instruments built the Explorer and later shrank it to silicon as the Explorer II and MicroExplorer; Xerox sold Interlisp-D machines. Symbolics was first to market and stayed ahead of TI, Xerox and LMI throughout.

The software layer was the shells and the consultancies — the "four horsemen": Teknowledge (formed 1981 by twenty founders, selling training, then the S.1 shell descended from EMYCIN, then custom applications, doing around $25 million a year in the 1984–86 period), IntelliCorp (KEE), Inference, and Carnegie Group. Alongside them sat MCC, the Austin research consortium that about twenty large American companies funded at $2–3 million each per year — $40–50 million annually, and, contrary to a persistent belief, no government money at all.

The demand was real, bounded, and partly subsidised, and all three of those matter. Real: XCON was in production and paying for itself. Bounded: Symbolics sold high-end development machines to high-end Lisp programmers, and there are only so many of those. Harvey Newquist, writing in AI Trends '88 while it was happening, put the ceiling in a sentence — "With an installed worldwide base of some 7000 LISP machines, there is even the possibility that there are actually more [LISP] machines in the marketplace than there are experienced LISP hackers. Frightening." Francis Feeney, Symbolics' assistant general counsel, said the same thing from the inside: "When we were growing so quickly here internally, I don't think anybody stopped and said, 'Well, wait a minute, once we saturate that fixed market, is there going to be a [broader] market for our system?'" And subsidised: Russell Noftsker, Symbolics' founding CEO, says many of the machines were bought by researchers funded through the Strategic Defense Initiative. SDI money did not go to Symbolics. It went to Symbolics' customers, who were building expert systems with it, and who — in the MIT case study's phrase — constituted "a relatively forgiving target market." DARPA's own Strategic Computing Initiative added roughly a billion dollars of federal computing research between 1983 and 1993, and had spent $100 million across 92 projects at 60 institutions by 1985.

Then three things happened at once.

The first was substitution. General-purpose workstations got good enough. A Symbolics 3600-series system in 1988 ran from $36,000 for a low-end configuration to $125,000 for a high-end one; a Sun could be had from around $14,000, ran Lisp well enough to develop in — via Lucid, Franz and the other portable implementations — and ran everything else too. Newquist again, contemporaneously: "Given the cost of individual LISP workstations, it was inevitable that users would look for alternate sources to deliver and develop their AI applications. General-purpose workstation vendors such as Sun Microsystems and Apollo computers have been quick to fill this bill." The difference, as the MIT case study puts it, "could be measured not in percent, but in hundreds of percent." David Moon, one of Symbolics' founders, described the end of the world the company was built for in one line: "the age of 80% gross margins in the computer industry was vanishing, and so was the age when selling 3000 machines a year was a big success."

The second was the withdrawal of the subsidy. Jacob T. ("Jack") Schwartz took over DARPA's Information Processing Techniques Office in 1987, dismissed expert systems as "clever programming," and cut AI funding "deeply and brutally," in McCorduck's account eviscerating the Strategic Computing Initiative. His stated principle was that DARPA should "surf" rather than "dog paddle" — fund the wave that was actually coming — and he did not think AI was it.

The third was disintermediation, and it is the least-told and most interesting. The tools companies were displaced by their own customers. Peter Friedland, on the CHM panel: "You saw companies like IntelliCorp lose market in essence to the people who had come to them originally for help in building systems … once it was important to them why would they rely on some dinky little company to do it for them, as opposed to doing it themselves?" The technique had transferred. The vendor was what became redundant.

What the collapse actually looked like. Symbolics peaked in fiscal 1986 — $101.6 million in product revenue, $12.6 million in services — and turned in fiscal 1987 to $82.1 million and $21.6 million, with a substantial net loss. In fiscal 1988 products fell again to $55.6 million. Services was the only line that grew throughout, from $12.6 million to $21.6 million to $25.5 million: the company was increasingly being paid to keep alive what it had already sold. Restructuring began in September 1986; Brian Sear was hired as COO in December 1986 and cut costs; the resulting fight with Noftsker ended in January 1988 with the board forcing out both men. LMI went bankrupt in 1987 with its next-generation K-Machine unshipped, was bought out of Chapter 11 as GigaMos, and collapsed again around 1988. TI and Xerox left the field. Lucid failed. Symbolics filed for Chapter 11 in January 1993. On the software side, Teknowledge fired the executive running its custom-applications group in 1987 and, Brown says, "that really crippled the custom applications group in ways that they didn't recover from very quickly"; it sold itself to American Cimflex by 1989. Carnegie Group, which never had venture money and was funded by corporates who wanted the technology, was eventually sold in 1998 for about $15 million against north of $30 million in revenue — roughly half times revenue, which Burt Grad, moderating, noted was "a little low" for a software company in the late 1990s. Mark Fox's own diagnosis was not that AI had failed but that "we were all over the place … We weren't a pure play. And that's reflected in the valuation."

The contemporaneous press knew what it was watching and reached for the same metaphor the field had coined for itself: the Boston Globe ran "AI Alley's Longest Winter" on 18 December 1988, and Forbes ran "Where Lisp Slipped" on 16 October 1989.

Why a reading would cite it

The occasion is standing rather than singular, and it is already in the project's window.

The last reading recorded, on 14 August 2026, Bank of America's estimate that Broadcom's off-balance-sheet AI chip-leasing vehicle could carry $370 billion of senior debt by mid-2029 at 20-gigawatt scale, against $29 billion of maximum backstop exposure actually disclosed in the 10-Q. In the same 48 hours it recorded Anthropic's preliminary Q2 revenue passing $11.5 billion with positive adjusted operating income, and OpenAI's enterprise business overtaking consumer at a $40 billion annualised run rate. Those two facts point in opposite directions about the same question — is there enough real demand under the capex — and it is exactly the question 1987 has a documented answer to, for one prior case.

The answer 1987 gives is not "the boom will bust." It is a shape, and the shape is more specific and more useful than the analogy it usually gets flattened into:

• The demand was real and the vendors still died. XCON was in production and saving money. DEC kept using it. The applications did not stop working when the machines that hosted them stopped selling. Nothing about a vendor collapse requires the underlying use case to have been fake, and a reading that treats a valuation event as a verdict on capability has skipped a step.
• What killed the hardware was a cheaper substrate that was good enough, not a failure of the application. This is the single most transferable finding. The Symbolics machine was better at its job and lost anyway, at 3–9× the price of something adequate that also did everything else. When a reading covers price compression — US frontier prices down roughly 25% in a month, open weights at Apache 2.0, a cheap-tier model taking a nine-point coding jump — the question 1987 says to ask is not "is the frontier still ahead" but "is the cheap thing now good enough for the work people are actually paying for."
• A concentrated customer base can be a ceiling nobody measured. Roughly 7,000 machines, possibly more machines than experienced Lisp programmers. The modern analogue is not obvious and should not be asserted, but the discipline is: when a reading meets a growth number, the 1987 question is what bounds it.
• Government-adjacent demand can be a subsidy that isn't labelled one. SDI money never appeared on Symbolics' books as government revenue; it arrived as customers. Sovereign and defence AI spending in 2026 has the same property.
• The suppliers can be disintermediated by their own customers. The IntelliCorp failure mode — the technique transfers, the buyer brings it in-house, the vendor becomes redundant — is a live question for every company currently selling AI tooling to firms that are hiring their own AI teams.
• The deflation was slow and legible only in filings. Six and a half years from the peak to Chapter 11 for the market leader; twenty-six to Teknowledge's bankruptcy. If a reading ever has to judge whether a downturn has started, this entry says that the answer arrives in fiscal-year revenue lines and not in a headline, and that the participants argued about management the whole way down.

There is a second, opposite reason to cite it, and it is the more common one in practice: as a check on the analogy itself. "AI winter" is the most-deployed historical comparison in AI commentary and it is nearly always deployed without its numbers. The entire corporate AI spend of 1985 was "over a billion dollars." One 2026 off-balance-sheet financing vehicle is projected at $370 billion of senior debt. Those differ by more than two orders of magnitude, and they differ in kind — venture equity and federal research grants then, senior debt and private credit now. A reading that reaches for 1987 to argue that the present boom will end the same way is asserting something the 1987 record cannot support. A reading that reaches for it to describe what a vendor-side deflation looks like from inside is on solid ground. This entry exists to make that distinction available, in both directions, and it is not evidence for either.

What it got right, and what it got wrong

Not required for moment. Included because there are dated, gradeable claims sitting inside this event, and the base rate this canon is assembling is worth more with them than without.

Right, and this is the entry's best single artifact — Minsky and Schank, made 1984, due 1987. At the AAAI annual meeting in 1984, Roger Schank and Marvin Minsky — both of whom had lived through the first winter — told the business community that enthusiasm had spiralled out of control and that disappointment would follow. They described a chain reaction by analogy with nuclear winter: pessimism in the AI community, then pessimism in the press, then a severe cutback in funding, then the end of serious research. Graded on the first three links: correct, and on a three-year horizon. The press turned (the Globe's "AI Alley's Longest Winter," December 1988). The funding was cut (Schwartz at IPTO, 1987). Graded on the fourth link: wrong, and importantly so. Serious research did not end. Backpropagation had been published in 1986, Pearl's work on belief networks was arriving in 1988, and both are direct ancestors of everything in a 2026 reading. The winter killed a market and a hardware architecture; it did not kill the field, and the two most consequential lines of work of the next forty years were being laid down while it ran. Anyone citing Minsky and Schank as vindicated should carry the fourth clause too.

Wrong — Symbolics' own three-year plan, made 1986, due 1989. The MIT case study reproduces two charts side by side: the company's 1986 forecast for fiscal 1987–89, and what actually happened. Read off the chart rather than given as figures, so approximate: the forecast has revenue climbing from about $114 million to roughly $137 million, $160 million and $190 million, with income rising from around $13 million to something like $20, $35 and $55 million. Actual: revenue fell to about $104 million and then about $81 million, and income went to roughly minus $25 million and then minus $35 million. A company at the top of its market, with better information about its own order book than anyone outside it, forecast about 67% revenue growth over three years and delivered a 29% decline into losses. That is the base rate this canon is for, and it is worth more than any outside commentator's prediction from the same year, because management had every instrument and still could not see one year ahead.

Right, and stated before the fact by people with everything to lose — expert systems would not scale. Doug Lenat, on the CHM panel, describing why he took the MCC job in the mid-1980s: "just at that point I'd been talking loudly, along with Alan Kay and several others, about what we saw as the impeding failure of expert systems to scale up for various reasons involving things like the difficulty of maintaining a consistent knowledge base. As the number of rules goes up, you don't want the time to enter 'rule n+1' to keep going up with n." Graded: correct. This is the same brittleness-and-maintenance diagnosis that mycin-1976 records Shortliffe and Shah drawing in retrospect, and it was being made in public before the market turned, by the people whose careers were inside it. Lenat's proposed fix, Cyc, is a fifty-year project whose grade is a separate question this entry does not settle.

Wrong — that machines would replace physicians within a decade. Tom Knight attributes to Edward Feigenbaum the hype that computers would be able to replace medical doctors within ten years; the Boston Globe ran "Computers Seen Replacing MDs, Lawyers" on 1 November 1984. Made mid-1980s, due mid-1990s. Graded: false, by a wide margin, and still open in 2026 — mycin-1976 traces the same claim through the 2026 Science study and the 2024 randomised trial that came back null. This is the clearest instance in the file of an overclaim by a senior figure inside the field being the thing that inflated the market that then deflated.

Wrong — that the AI hardware industry was replaced in a single year. Made by Crevier in 1993 and repeated ever since; it is the source of the proposals line this entry corrects. Graded: false as stated, true in spirit about the category. Nothing was replaced in a single year. The peak-to-bankruptcy interval for the market leader was six and a half years.

Right, and it cuts the other way — the technique paid for the whole programme during the winter. DART, the Dynamic Analysis and Replanning Tool, went into use at USTRANSCOM as a prototype in 1991 for Operation Desert Storm. DARPA's own verdict, attributed to then-director Victor Reis, is that DART repaid the agency's entire thirty years of AI investment in a matter of months. Made and due within the winter itself. A reading that treats 1987 as proof that the money was wasted has to answer this, and it comes from the funder rather than the field.

Unresolved rather than graded — whether the Lisp machines were doomed independently. Patrick Winston, then director of the MIT AI Lab, says Symbolics was "doomed" from its conception because it ignored the delivery market, and "would have hit a wall whether the AI industry had crashed or not." The MIT study adds late delivery (the 3600 about a year late, the I-Machines about two), incompatibility, and manufacturing separated from engineering by a continent for reasons its authors call unjustifiable. If Winston is right, then the market collapse was the occasion rather than the cause, and 1987 is a worse analogy for a sector-wide event than it looks. This entry does not settle that, and a reading leaning hard on the analogy should know it is contested by the person who ran the lab the company came out of.

Commonly misused as

Not required for moment. Included because "AI winter" is, by a distance, the most frequently misused historical claim in AI commentary, and this entry's most useful service is likely to be stopping a lazy use of it rather than enabling one.

• "AI winters are cyclical, so another one is due." The load-bearing misuse. Two contractions roughly fifteen years apart is not a cycle: there is no established period, no isolated mechanism, and no statement of the conditions under which it recurs. The word "winter" does the work here — it imports seasonality into an argument that has not earned it. What the record establishes is that one specific market, whose demand was bounded by a scarce skill and partly subsidised by one defence programme, contracted when a cheaper substitute arrived and the subsidy stopped. That is a mechanism, and a reading can check whether its parts are present. "It's been a while" is not.
• "It happened before at this scale, so it will happen again at this scale." The numbers do not survive contact. Total corporate AI spending in 1985: over a billion dollars. A single 2026 financing vehicle: $370 billion of projected senior debt. The 1987 case tells you about the shape of a vendor-side deflation. It tells you nothing about the magnitude or probability of a 2020s one, and the composition is different in a way that matters — equity and research grants absorb losses differently from senior debt.
• "It was caused by the October 1987 crash." A conflation that appears often enough to be worth killing on the date alone. Black Monday was 19 October 1987. Symbolics' fiscal year ended 30 June 1987, already in loss, after a restructuring that had begun in September 1986. The AI market had turned before the stock market did, for reasons — Sun workstations, saturated customer base, SDI wind-down — that have nothing to do with portfolio insurance.
• "The technology was a dead end." It was absorbed, which looks identical to death if you only track vendors. Rule engines did not disappear; they became infrastructure, under a different name, in the business-rules-management industry (ILOG, bought by IBM in 2009; FICO; Pegasystems) and in the openly available CLIPS lineage. mycin-1976 records the sharper version of this from Shortliffe and Shah in 2022: the "knowledge is power" aphorism "has been somewhat forgotten in today's AI research and application communities—arguably to their detriment." That is a position a reading may cite as a position, not as a settled finding.
• "Expert systems never worked." XCON ran in production for years and saved DEC real money; DART repaid DARPA's whole AI programme. What failed was the economics of selling them: knowledge acquisition was human-intensive and stayed that way, the incremental cost per system did not fall, maintenance scaled badly with rule count, and — Friedland's point — the buyers eventually did it themselves. "Expensive to build and maintain, and impossible to sell at a margin" is the accurate charge, and it is a different and more interesting one than "didn't work."
• "DARPA pulled out of AI in 1987." Schwartz cut deeply and said something quotable, but Strategic Computing continued, some projects survived, and DART came out the other end of it. Funding priority shifted; the agency did not leave.
• "The 1980s bust proves today's AI companies are overvalued." It proves nothing about today's companies. It is one prior case, in a market roughly two orders of magnitude smaller, with a different financing structure, a different cost curve, and a customer base bounded by a scarce human skill that has no clean modern equivalent. Nothing in this file is evidence, nothing in it deposits into the ledger, and nothing in it moves the needle. The entry supplies a documented shape and the questions that go with it. It does not supply a forecast, and a reading that borrows its authority to make one is doing the thing rule 7 exists to prevent.

Sources

Read directly, in full or in the parts named. Alvin Graylin, Kari Anne Hoier Kjolaas, Jonathan Loflin and Jimmie D. Walker III, "Symbolics, Inc.: A failure of heterogeneous engineering," a term paper for MIT 6.933J (The Structure of Engineering Revolutions), posted on MIT OpenCourseWare — read pages 1–14 and 21–34 of 34; the intervening technical section on Lisp machine hardware design was skipped. This is the source of the fiscal 1986/87/88 product and services revenue figures, the two financial charts, the September 1986 restructuring, the Sear hiring and the January 1988 board fight, the 1988 price ranges, the Newquist, Feeney, Moon, Winston and Sear quotations, the SDI-as-indirect-market argument, and the product-delay figures. Its own sources are Symbolics annual reports, interviews with Russell Noftsker (founding CEO), Thomas Knight and Harvey Newquist, email from David A. Moon, and contemporaneous trade press — Business Week (13 June 1988), Electronic Business (1 September 1988), New England Business (1988), Industry Week (16 November 1987), and the Boston Globe and Forbes pieces named above. It is a student paper and should be weighted as one; its strength is the primary interviews, and it contains at least one obvious typo (the January 1988 board fight is printed as 1998).

Computer History Museum, "AI: Expert Systems Pioneer Meeting, Session 4: Business Reviews of Initial Companies," recorded 14 May 2018 in Mountain View, moderated by Burt Grad and David C. Brock, participants including Mark Fox, Denny Brown, Peter Friedland, Doug Lenat and Ed Feigenbaum — read pages 26–36 of 39. Source of the Carnegie Group figures and sale, the Teknowledge history including the 1987 firing, the 1989 American Cimflex sale and the December 2013 bankruptcy, the EMYCIN→KS300→S.1 lineage that justifies descends_from, Friedland's disintermediation observation, and Lenat's account of MCC and of warning about scaling before the bust. The document's cover page gives its reference number as X8652.2019 and its page footers give X8903.2019; the discrepancy is in the original.

Wikipedia, "AI winter," read as wikitext for its citations rather than its prose. Source of the Minsky/Schank AAAI 1984 account, the "half a billion dollars replaced in a single year" claim, the XCON $40-million-over-six-years figure, the Schwartz quotations, and the SCI figures. The underlying books were not read and are named here so a reading can go to them: Daniel Crevier, AI: The Tumultuous History of the Search for Artificial Intelligence (1993), pp. 203 and 209–210; Harvey Newquist, The Brain Makers (1994), pp. 189–201; Pamela McCorduck, Machines Who Think (2nd ed., 2004), pp. 426–431. Note that Wikipedia renders the DARPA official's surname "Schwarz"; the person is the NYU mathematician Jacob T. Schwartz, who headed IPTO from 1987.

Consulted through search-result and fetch summaries rather than full text, and flagged as weaker accordingly: XCON's rule count (~2,500), order volume (80,000 by 1986), the alternative $25-million-a-year savings figure, and the eight-person maintenance team; LMI's 1987 bankruptcy, the unshipped K-Machine, and the GigaMos reorganisation; Symbolics' Chapter 11 filing in January 1993; Schwartz's IPTO tenure of 1987–1989; DART's 1991 Desert Storm deployment and the Victor Reis payback claim; Alex Roland and Philip Shiman, Strategic Computing: DARPA and the Quest for Machine Intelligence, 1983–1993 (MIT Press, 2002) for the billion-dollar decade figure; the roughly $500 million cost of Japan's Fifth Generation project; and the tfeb.org essay "The lost cause of the Lisp machines" (18 November 2025), by a programmer who used both families of Lisp machine professionally from 1989, for the argument that commodity RISC hardware had already overtaken them on benchmarks by 1987 — a useful corrective to the nostalgia, read only in summary.

Attempted and failed. Communications of the ACM, "How the AI Boom Went Bust," returned HTTP 403. Curt Monash's "AI memories — expert systems" on softwarememories.com failed DNS resolution. Both looked directly relevant and neither was read. A widely repeated claim that Symbolics stock fell "from $45 to under $5 in months" circulates only in low-quality secondary write-ups; I could not trace it to a filing or a contemporaneous report, it is named here as unverified, and nothing above depends on it. The precise fiscal-1987 net loss of $25.5 million likewise comes from a secondary summary rather than a filing, and happens to coincide with a services-revenue figure in the MIT paper, which is grounds for suspecting a transcription collision; what the paper's own chart independently supports is a fiscal-1987 loss of roughly $25 million, and that is the figure this entry relies on.

The 2026 citation occasion — the Broadcom financing estimate of 14 August 2026, the Anthropic and OpenAI revenue disclosures of 14–15 August 2026, and the frontier price movements — is as recorded in this project's own digests/2026-08-15-12.md, which holds the primary links. It is named here as an occasion to cite this entry, not as evidence for anything in it. Nothing in this file is evidence, nothing in it is deposited in the ledger, and nothing in it touches the needle.