In August 2024 the reading was that the AI shakeout would not look like the dot-com bust. Eighteen months later, 72 percent of the logos were gone and almost nobody noticed.
Yesterday the Financial Times reported that tech stocks fell after the biggest AI companies called for a slowdown in the technology’s development. Memory and chip makers took the steepest losses. SoftBank, which holds around 13 percent of OpenAI, dropped nearly 11 percent. Ten-year Treasury yields topped 5 percent and Brent crude approached $110.
A lot of people will read that as the beginning of something.
I want to suggest it is closer to the middle, and that the part worth worrying about has already happened.
What I said in August 2024
Two years ago, in a public exchange on LinkedIn, Bill DeMartino and I disagreed about exactly this. His position was that there would be a shakeout, but nothing near the impact of the dot-com bust. What I put in front of him was not a forecast. It was a reading of a pattern the archive had already recorded twice.
I told him he was right about the first part. It was not going to be a spectacular fireworks display.
I would describe the impact more as high blood pressure — there is a reason why they call it the “silent killer.” The immediate fallout will be the disappearing logos from the solution provider maps. You can call that a little fireworks.
However, the damage to supply chains now will be felt for the next couple of years because we will have to make up for the lost efficiencies and people talent that hasn’t been developed as expected. And I am not talking about AI or GenAI talent development. I am talking about agent-based problem-solving talent development.
I also described a contraction of the solution provider map back toward where it had been in 2019, and said that for procurement specifically it would end up more devastating than 2001 — not despite being quieter, but because of it.
That post is still there, dated 24 August 2024.
What the measurement showed
In February 2026 we stopped discussing it and counted.
ARA™ RAM 2025™ was run against the archive and the current solution map — not to forecast anything, but to verify what had already been written against what had actually happened to the logos.
Using a four-state outcome classification — Defunct, Absorbed, Merged or Renamed, Still Operating Independently — 72 percent of the ProcureTech vendor logos on the baseline solution map no longer exist as independent entities. Not “might not.” Do not.
Twenty-two percent defunct. Thirty-one percent absorbed with the brand discontinued and the roadmap subordinated. Nineteen percent merged or renamed with the product identity uncertain. Twenty-eight percent still operating as they were.
Then we went further back. Of the ten solution providers I profiled as the “10 for 2010,” exactly one — Ivalua — is still thriving independently. Nine out of ten, across a full technology cycle.
The full breakdown is here.
Two measurements, fifteen years apart. And here is the part that matters for this week: neither of them required a crash. No index fell 0.4 percent on the day the 2010 cohort disappeared. There was no headline. The logos went quietly, one acquisition and one wind-down at a time, over years.
Two lines, thirty years
The vendor-count curve is illustrative of the pattern rather than a measured series; the 72 percent and the 2010 cohort are measured. The flat line is the one to watch.
The green line is the one everybody watches. It rises through a hype cycle, peaks, and falls. It did it in 2000 and it is doing it now.
The gold line is the one nobody plots. The proportion of technology initiatives that fail to achieve their expected results has not meaningfully moved through any of it — not through the burst, not through the recovery, not through ERP, SOA, cloud, SaaS, analytics or AI.
Look at what that means. The vendor population collapsed by roughly 80 percent after 2001 and the failure rate did not improve. The population more than doubled between 2016 and 2023 and the failure rate did not improve either. The line that moves has no observable effect on the line that matters.
That is not a coincidence to be explained away. It is evidence about where the determining conditions actually live, and they do not live in the vendor population.
Why quiet is worse than loud
The dot-com bust was loud, visible and dated. Commerce One, worth billions at its peak, filed for bankruptcy in 2004. FreeMarkets, which genuinely invented something, was absorbed by Ariba the same year. Ariba itself survived, diminished, and was acquired by SAP in 2012.
None of those were frauds. The technology worked. The people were credible. CPOs who selected them made defensible decisions on the best information available — and then inherited roadmaps they had never chosen.
But everyone knew it happened. There was a date. Boards asked questions. Post-mortems were written, and the organizations that were affected at least understood what had happened to them.
The contraction I described in 2024 has no date. There is no morning where a CPO wakes up to the news that their vendor is gone — there is a rebrand, a “strategic combination,” a roadmap that quietly stops moving, an account manager who is suddenly covering four times as many clients. By the time it is obvious, it is eighteen months old and nobody is asking questions about it, because nothing happened on any particular Tuesday.
High blood pressure does not hurt. That is the entire problem with it.
The cost nobody is counting
The logos are the measurable part, and 72 percent is a large number. But it is the small half of the damage.
Here is what the last two years actually cost, and none of it appears on an index.
Every organization that bought capability in place of building it has an operating model it still does not understand. Every practitioner who spent 2024 and 2025 evaluating platforms spent those years not developing agent-based problem-solving skill. Every team that was told the tool would surface the answer has less practice at finding one than it had before.
That is what I meant in 2024 by talent that has not been developed as expected. Not model literacy, which is being trained everywhere — the ability to reason toward an outcome when the system has not handed you one.
And it explains the gold line. Each era arrives, capability increases enormously, organizations buy the capability instead of building the readiness to use it, and the outcome does not move. Then the vendors contract, the next era arrives, and the cycle runs again with different logos.
What a slowdown actually means for ProcureTech
There is a more immediate question buried in yesterday’s story, and it is the one a practitioner should be asking.
If the largest AI companies genuinely slow down, what happens to the solution providers selling AI capability on top of them?
Much of what is currently marketed as AI in ProcureTech depends materially on foundation models the application vendor does not own or control. It rents the capability it is selling. That arrangement has three consequences, and none of them appear in a product demonstration.
The roadmap was never theirs to promise. A vendor that has sold you agentic capability arriving next year is relaying a schedule set by a company it has no contractual relationship with. If that schedule slips, nobody announces it. The feature simply keeps not shipping, and the explanation arrives as a quarterly release note rather than a conversation.
The cost floor moves underneath them. Inference pricing is the one input a wrapper cannot control. Expensive capital and rising rates at the foundation layer eventually surface as a per-call price, and the vendor’s margin absorbs it until it cannot. That is a solvency question, and it belongs to the buyer the moment the contract is signed rather than before.
Consolidation accelerates. It does not slow. This is the part that runs against intuition. A slowdown does not protect the smaller providers — it starves them. Rapid capability improvement is what let a small player look competitive with a large one. If the underlying models stop improving quickly, the differentiator reverts to distribution and balance sheet, and those belong to the incumbents.
Which means a slowdown does not lower the 72 percent. It raises it.
And none of the three is visible in the software you evaluated, or written into the contract you signed.
What this means for a selection decision
If 72 percent of the logos on a solution map have already ceased to exist as independent entities, in ordinary weather, with no crash required, then a market-wide repricing is not a new risk. It is the same risk with the clock sped up.
Which makes the question I put in February the operative one this week:
The most important question is not which vendor should I select? It is can my organization absorb and sustain value from this investment if the vendor’s identity changes?
That is a readiness question. It is Phase 0™. And it is the question solution maps, Magic Quadrants and analyst rankings do not ask, because asking it would mean telling some clients not to proceed.
Vendor survivability is not a selection variable. It is a readiness variable, and it always was.
One honest caveat
A single day’s market move is not evidence of anything. Yesterday’s selloff is a trigger for the argument, not proof of it — the evidence that matters is what rating agencies do, what the filings show, and which logos are still independent in eighteen months.
I am not forecasting a crash. What was written in 2024 was that it would not arrive as one, and nothing in the evidence since has changed that.
What I am saying is that the damage does not require a crash to occur. It has been occurring, quietly, in ordinary weather, for two years — and it will keep occurring whether or not the fireworks ever go off.
The fireworks are not the story.
They never were.
-30-
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
Jon W. Hansen, FCIPS — Procurement Insights | Hansen Models™
The same readiness problem appears one layer down — in how you work with the models the wrappers rent. I am running a free 30-minute lab on that smaller version of it: how to select, work with and progressively engage AI models so the output gets more accurate rather than just more polished. It is a working session, not a presentation — you make the calls before you see what I did.
Friday, 25 September, 9:30 AM ET. Everyone who attends gets a copy of Thinking With the Machine.
Register: https://www.linkedin.com/events/7504567838724997120/
Related
“Tech Stocks Fall After Big AI Groups Call for Slowdown.” The Fireworks Are Not the Story. They Never Were.
Posted on September 14, 2026
0
In August 2024 the reading was that the AI shakeout would not look like the dot-com bust. Eighteen months later, 72 percent of the logos were gone and almost nobody noticed.
Yesterday the Financial Times reported that tech stocks fell after the biggest AI companies called for a slowdown in the technology’s development. Memory and chip makers took the steepest losses. SoftBank, which holds around 13 percent of OpenAI, dropped nearly 11 percent. Ten-year Treasury yields topped 5 percent and Brent crude approached $110.
A lot of people will read that as the beginning of something.
I want to suggest it is closer to the middle, and that the part worth worrying about has already happened.
What I said in August 2024
Two years ago, in a public exchange on LinkedIn, Bill DeMartino and I disagreed about exactly this. His position was that there would be a shakeout, but nothing near the impact of the dot-com bust. What I put in front of him was not a forecast. It was a reading of a pattern the archive had already recorded twice.
I told him he was right about the first part. It was not going to be a spectacular fireworks display.
I also described a contraction of the solution provider map back toward where it had been in 2019, and said that for procurement specifically it would end up more devastating than 2001 — not despite being quieter, but because of it.
That post is still there, dated 24 August 2024.
What the measurement showed
In February 2026 we stopped discussing it and counted.
ARA™ RAM 2025™ was run against the archive and the current solution map — not to forecast anything, but to verify what had already been written against what had actually happened to the logos.
Using a four-state outcome classification — Defunct, Absorbed, Merged or Renamed, Still Operating Independently — 72 percent of the ProcureTech vendor logos on the baseline solution map no longer exist as independent entities. Not “might not.” Do not.
Twenty-two percent defunct. Thirty-one percent absorbed with the brand discontinued and the roadmap subordinated. Nineteen percent merged or renamed with the product identity uncertain. Twenty-eight percent still operating as they were.
Then we went further back. Of the ten solution providers I profiled as the “10 for 2010,” exactly one — Ivalua — is still thriving independently. Nine out of ten, across a full technology cycle.
The full breakdown is here.
Two measurements, fifteen years apart. And here is the part that matters for this week: neither of them required a crash. No index fell 0.4 percent on the day the 2010 cohort disappeared. There was no headline. The logos went quietly, one acquisition and one wind-down at a time, over years.
Two lines, thirty years
The vendor-count curve is illustrative of the pattern rather than a measured series; the 72 percent and the 2010 cohort are measured. The flat line is the one to watch.
The green line is the one everybody watches. It rises through a hype cycle, peaks, and falls. It did it in 2000 and it is doing it now.
The gold line is the one nobody plots. The proportion of technology initiatives that fail to achieve their expected results has not meaningfully moved through any of it — not through the burst, not through the recovery, not through ERP, SOA, cloud, SaaS, analytics or AI.
Look at what that means. The vendor population collapsed by roughly 80 percent after 2001 and the failure rate did not improve. The population more than doubled between 2016 and 2023 and the failure rate did not improve either. The line that moves has no observable effect on the line that matters.
That is not a coincidence to be explained away. It is evidence about where the determining conditions actually live, and they do not live in the vendor population.
Why quiet is worse than loud
The dot-com bust was loud, visible and dated. Commerce One, worth billions at its peak, filed for bankruptcy in 2004. FreeMarkets, which genuinely invented something, was absorbed by Ariba the same year. Ariba itself survived, diminished, and was acquired by SAP in 2012.
None of those were frauds. The technology worked. The people were credible. CPOs who selected them made defensible decisions on the best information available — and then inherited roadmaps they had never chosen.
But everyone knew it happened. There was a date. Boards asked questions. Post-mortems were written, and the organizations that were affected at least understood what had happened to them.
The contraction I described in 2024 has no date. There is no morning where a CPO wakes up to the news that their vendor is gone — there is a rebrand, a “strategic combination,” a roadmap that quietly stops moving, an account manager who is suddenly covering four times as many clients. By the time it is obvious, it is eighteen months old and nobody is asking questions about it, because nothing happened on any particular Tuesday.
High blood pressure does not hurt. That is the entire problem with it.
The cost nobody is counting
The logos are the measurable part, and 72 percent is a large number. But it is the small half of the damage.
Here is what the last two years actually cost, and none of it appears on an index.
Every organization that bought capability in place of building it has an operating model it still does not understand. Every practitioner who spent 2024 and 2025 evaluating platforms spent those years not developing agent-based problem-solving skill. Every team that was told the tool would surface the answer has less practice at finding one than it had before.
That is what I meant in 2024 by talent that has not been developed as expected. Not model literacy, which is being trained everywhere — the ability to reason toward an outcome when the system has not handed you one.
And it explains the gold line. Each era arrives, capability increases enormously, organizations buy the capability instead of building the readiness to use it, and the outcome does not move. Then the vendors contract, the next era arrives, and the cycle runs again with different logos.
What a slowdown actually means for ProcureTech
There is a more immediate question buried in yesterday’s story, and it is the one a practitioner should be asking.
If the largest AI companies genuinely slow down, what happens to the solution providers selling AI capability on top of them?
Much of what is currently marketed as AI in ProcureTech depends materially on foundation models the application vendor does not own or control. It rents the capability it is selling. That arrangement has three consequences, and none of them appear in a product demonstration.
The roadmap was never theirs to promise. A vendor that has sold you agentic capability arriving next year is relaying a schedule set by a company it has no contractual relationship with. If that schedule slips, nobody announces it. The feature simply keeps not shipping, and the explanation arrives as a quarterly release note rather than a conversation.
The cost floor moves underneath them. Inference pricing is the one input a wrapper cannot control. Expensive capital and rising rates at the foundation layer eventually surface as a per-call price, and the vendor’s margin absorbs it until it cannot. That is a solvency question, and it belongs to the buyer the moment the contract is signed rather than before.
Consolidation accelerates. It does not slow. This is the part that runs against intuition. A slowdown does not protect the smaller providers — it starves them. Rapid capability improvement is what let a small player look competitive with a large one. If the underlying models stop improving quickly, the differentiator reverts to distribution and balance sheet, and those belong to the incumbents.
Which means a slowdown does not lower the 72 percent. It raises it.
And none of the three is visible in the software you evaluated, or written into the contract you signed.
What this means for a selection decision
If 72 percent of the logos on a solution map have already ceased to exist as independent entities, in ordinary weather, with no crash required, then a market-wide repricing is not a new risk. It is the same risk with the clock sped up.
Which makes the question I put in February the operative one this week:
The most important question is not which vendor should I select? It is can my organization absorb and sustain value from this investment if the vendor’s identity changes?
That is a readiness question. It is Phase 0™. And it is the question solution maps, Magic Quadrants and analyst rankings do not ask, because asking it would mean telling some clients not to proceed.
Vendor survivability is not a selection variable. It is a readiness variable, and it always was.
One honest caveat
A single day’s market move is not evidence of anything. Yesterday’s selloff is a trigger for the argument, not proof of it — the evidence that matters is what rating agencies do, what the filings show, and which logos are still independent in eighteen months.
I am not forecasting a crash. What was written in 2024 was that it would not arrive as one, and nothing in the evidence since has changed that.
What I am saying is that the damage does not require a crash to occur. It has been occurring, quietly, in ordinary weather, for two years — and it will keep occurring whether or not the fireworks ever go off.
The fireworks are not the story.
They never were.
-30-
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
Jon W. Hansen, FCIPS — Procurement Insights | Hansen Models™
The same readiness problem appears one layer down — in how you work with the models the wrappers rent. I am running a free 30-minute lab on that smaller version of it: how to select, work with and progressively engage AI models so the output gets more accurate rather than just more polished. It is a working session, not a presentation — you make the calls before you see what I did.
Friday, 25 September, 9:30 AM ET. Everyone who attends gets a copy of Thinking With the Machine.
Register: https://www.linkedin.com/events/7504567838724997120/
Share this:
Like this:
Related