A New Beginning

Why the AI Correction Makes This the Moment for PSL/i

Glenn Hutchinson | Architect of the Personal Semantic Layer | Hutchinson & Co. | July 2026

For three years the entire industry has been making one bet, over and over, in larger and larger amounts. The bet is that if you make the model bigger and give it more compute, it becomes more useful. That bet built the largest capital cycle in the history of technology. It is now running into a wall that more money cannot move, and the shape of that wall is the whole reason PSL/i should exist right now instead of later. I want to start with the numbers, because the numbers are no longer arguable.

The money went somewhere, and it is not coming back the way people promised.

In 2026 the five largest hyperscalers are on track to spend more than $700 billion in capital expenditure, nearly double what they spent in 2025 and roughly triple what they spent in 2024. Of that amount, close to $450 billion is aimed directly at AI infrastructure: GPU clusters, data centers, and the power required to run them.¹ Bank of America and Evercore expect 2027 to cross a trillion.¹

Now hold that against what is coming back. Sequoia’s David Cahn has put the gap between infrastructure spending and actual AI ecosystem revenue at roughly $600 billion a year.² Bain & Company framed the same problem from the other direction, working out that to finance the infrastructure the industry says it needs by 2030, AI companies would have to generate around $2 trillion in annual revenue, and on current trajectories they fall short by roughly $800 billion.³ Allianz Research measured the divergence between capex and revenue at 46 percent, wider than the 32 percent divergence that preceded the 2001 telecom collapse.²

This is not a rumor of trouble; it is the arithmetic of the people building it. The market has started to notice, and when Meta raised its capex guidance, its stock fell more than nine percent in a single session, which is the market saying out loud that it no longer treats more spending as automatically good news.²

The frontier labs themselves tell the same story from the inside. OpenAI reportedly lost $13.5 billion in the first half of 2025 alone, and by one detailed accounting, it spends about two dollars for every one dollar of inference revenue, before you count research, sales, marketing, or real estate.⁴ Sam Altman has acknowledged that the $200-a-month Pro subscription operates at a loss.⁴ Anthropic, running a leaner model business, still projects a $14 billion loss in 2026 and does not expect positive free cash flow before 2028.⁴ These are the strongest companies in the field, and this is what winning currently looks like.

The wall is not a money problem. It is an architecture problem.

Here is the part that matters for anyone deciding where to put capital next. All of that spending is buying a better version of the thing that already does not hold up at the point of use.

In August 2025 MIT’s NANDA initiative published The GenAI Divide: State of AI in Business 2025, built from 150 leadership interviews, a 350-person survey, and 300 documented deployments. The headline finding was that roughly 95 percent of enterprise generative-AI pilots deliver no measurable impact on the profit-and-loss statement, and that number holds against a reported $40 billion spent on those internal efforts.⁵

The reason the study gives is the one the whole capital cycle keeps skipping past. The pilots do not fail because the models are not smart enough; they fail because generic systems cannot learn or adapt to a specific organization’s context and workflow.⁵ The intelligence is capable enough, and the foundation underneath it is missing.

You can watch the same failure at the smaller scale of a single agent. Production agents tend to hold up for a week or two and then start to drift, contradicting decisions they made earlier, re-fetching data they already have, and ignoring constraints they were given at the start, even though neither the model nor the prompt has changed.⁶ The reason is structural: the context window is being used as if it were memory, and it is not memory. It is volatile, it disappears when the session ends, and it degrades under load. One 2026 study measured constraint compliance falling from 73 percent at the fifth turn of a conversation to 33 percent by the sixteenth.⁶ The system forgets what it was told in the middle of doing the work it was told to do.

I have written before that current AI does not just forget you between sessions, it starts losing meaning inside a single conversation. The industry’s answer has been what people call “good enough”: better prompts, tighter retrieval, and faster iteration, all of it aimed at getting the output close enough that the amnesia underneath does not show. For a quick answer, good enough is genuinely good enough. For anything you are trying to build on, it falls apart, and $450 billion a year is now being spent on making the wrong layer faster.

The industry has finally noticed, and is building the layer on top instead of underneath.

Give the field credit, it has stopped pretending the amnesia does not exist. The most discussed enterprise-AI thesis of the past year, Foundation Capital’s case for a “context graph,” is an open admission that systems of record capture what happened but never why, and that the reasoning behind a decision is the thing worth owning.⁷ Around that idea, a whole vocabulary has formed. Teams now talk about knowledge graphs, semantic layers, and memory layers as the missing piece, and capital is beginning to move toward all three.

The trouble is not the diagnosis; it is that every one of these answers is a bolt-on. Each one sits on top of the model or the data warehouse and tries to hand meaning back at the moment of the query, yet none is both deterministic and persistent, which are the two properties that would make it a foundation instead of a patch.

Take knowledge graphs and GraphRAG, the most technically serious attempt. Retrieval there is probabilistic by construction, so the system fetches what it estimates is relevant, usually in a single pass, and the same question can surface different information depending on how the graph was built and how it was walked. The research literature is now naming retrieval drift and retrieval hallucinations directly as problems created by imperfect knowledge graphs.⁸ A graph can hold every fact required for a four-step answer and still miss it, because single-pass retrieval never walks the chain.⁹ That is not a foundation, it is a smarter lookup that is right most of the time, which is precisely the good enough that collapses the moment you try to build on it.

The term “semantic layer” has been adopted too, but for something far narrower than it sounds. As the market now uses it, in tools like Cube, dbt, and AtScale, a semantic layer governs analytics. It defines what “revenue” means, which joins are permitted, and who is allowed to query what, so that an agent pulls from certified metrics instead of writing raw SQL against the warehouse.¹⁰ That is genuinely useful, but it is not what the term should mean. It holds metric definitions for a database, not who you are, what you have decided, or why. It is governance for queries, not a persistent foundation for a life or a business.

Then there is the memory layer, and here the money is already explicit. Mem0 raised $24 million to build a portable “memory layer for AI apps,” and the demand is not in doubt, with tens of millions of package downloads and API calls that climbed from 35 million to 186 million inside a single year.¹¹ The demand is real, and the architecture is the problem. This kind of memory is retrieval over stored snippets, wired in one application at a time, and its own founder points out that the large labs have every incentive to keep memory locked behind their walls as a moat.¹¹ It is meaning rented back to you one app at a time, and it is neither deterministic nor truly yours.

So the pattern simply repeats one layer up. Having spent the first wave of capital making the model bigger, the industry is now spending the next wave on bolt-on ways to feed that model context, whether graphs, semantic layers, or memory services, each expensive to build and maintain, while none solves the actual problem, which is that meaning has to be held once, deterministically and persistently, and owned by the person or the entity it belongs to. The money has moved to the right neighborhood and is still buying the wrong thing.

Why this is the opening, not the obituary.

The easy read of these numbers is that AI is a bubble and the smart move is to wait. I think that read is exactly backwards, and the reason is where the money is misallocated.

The spending is not wrong because AI is worthless. It is wrong because it is concentrated in one layer, the model and the compute beneath it, while the layer that would actually make the model useful to a specific person or business does not exist yet. The scale of that $600 billion gap is evidence that model scaling is hitting diminishing returns at the point where meaning is supposed to live. That is not a case against AI, it is the clearest case anyone could make for the missing layer.

PSL/i is that layer, and it is defined by the two properties the bolt-ons never had. It is deterministic, so meaning is held once and returns the same each time, rather than being re-estimated by a probabilistic retrieval pass that can drift from one query to the next. It is also persistent, holding who you are, what you know, how you work, who you work with, what you have decided and why, across every session and every tool, owned by the individual rather than rebuilt from scratch each time a model is asked to do something. It does not sit on top of the context problem the way a knowledge graph, an analytics semantic layer, or a rented memory service does; it sits underneath all of them. When an agent finally acts, it acts on something real instead of on whatever survived sixteen turns of a decaying context window.

For an investor, the timing argument has never been stronger, precisely because of the correction rather than in spite of it. When capital believed bigger models would solve everything, a semantic foundation sounded like a nice-to-have beneath an infinite budget. Now the budget is finite, the returns on scale are visibly bending, and 95 percent of enterprise pilots are failing for the exact reason PSL/i addresses. For the first time, the market is motivated to look one layer down, and the capital already moving into context graphs, semantic layers, and memory startups is that same market reaching for this layer with the wrong tools. That reaching validates the need while leaving the deterministic, persistent version of it unbuilt. The thesis that the model is the product is repricing in real time, and what replaces it is the layer that makes any model worth deploying.

It’s not the end of the AI story, rather it’s the map for a new beginning. It points straight at the part nobody has built yet, the layer underneath the models where meaning finally holds still and belongs to the person it describes.

The companion papers and architecture documents supporting this work, including the Personal Semantic Layer framework and the Meaning Before Applications thesis, are available on my LinkedIn profile.

Sources

¹ Vin Vashishta, “$700 Billion in Capex. $50 Billion in Revenue. AI’s Math Is Broken,” Substack, 2026. https://vinvashishta.substack.com/p/700-billion-in-capex-50-billion-in. Jason Kirsch, “AI Spending Is Surging Faster Than Revenue and Markets Are Repricing,” Forbes, June 2, 2026. https://www.forbes.com/sites/jasonkirsch/2026/06/02/the-ai-capex-to-revenue-gap-is-widening—and-markets-are-starting-to-notice/

² Jason Kirsch, “AI Spending Is Surging Faster Than Revenue and Markets Are Repricing,” Forbes, June 2, 2026. Includes David Cahn and Sequoia’s estimate of the roughly $600 billion revenue gap, Allianz Research’s comparison of the 46 percent capex-to-revenue divergence with the 32 percent telecom divergence, and Meta’s single-session stock decline. https://www.forbes.com/sites/jasonkirsch/2026/06/02/the-ai-capex-to-revenue-gap-is-widening—and-markets-are-starting-to-notice/

³ Calcalist / CTech, “The Circular Economy of AI: How Big Tech Is Financing Itself,” 2026. Includes Bain & Company’s estimate that approximately $2 trillion in annual revenue would be needed by 2030 and the projected shortfall of roughly $800 billion, along with the approximately $40 billion invested in internal AI agents. https://www.calcalistech.com/ctechnews/article/z4lxiqbtw

⁴ Vin Vashishta, “$700 Billion in Capex. $50 Billion in Revenue. AI’s Math Is Broken,” Substack, 2026. Includes OpenAI’s reported $13.5 billion loss during the first half of 2025, the estimate of approximately two dollars in cost for every dollar of inference revenue, the Pro subscription operating at a loss, Anthropic’s projected $14 billion loss in 2026, and its expectation that positive free cash flow will not arrive before 2028. https://vinvashishta.substack.com/p/700-billion-in-capex-50-billion-in

⁵ MIT NANDA, The GenAI Divide: State of AI in Business 2025, reported by Fortune, August 18, 2025. The report found that approximately 95 percent of enterprise generative-AI pilots showed no measurable profit-and-loss impact, with failure driven primarily by integration, organizational context, and workflow adaptation rather than model quality. https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/

⁶ Mem0, “Context Window Is RAM, Not Storage: Why Most Agent Failures Happen and How to Fix Them in 2026.” Includes the characterization of the context window as volatile, non-persistent memory, the decline in constraint compliance from 73 percent at turn five to 33 percent at turn sixteen, and the “lost in the middle” degradation problem. https://mem0.ai/blog/context-window-is-ram-not-storage-why-most-agent-failures-happen-how-to-fix-them-in-2026

⁷ Jaya Gupta and Ashu Garg, “AI’s Trillion-Dollar Opportunity: Context Graphs,” Foundation Capital, December 22, 2025. The essay argues that existing systems of record capture outcomes without preserving the decision traces that explain how and why those outcomes were reached, and that context graphs built from those traces could become a foundational layer for enterprise AI. https://foundationcapital.com/ideas/context-graphs-ais-trillion-dollar-opportunity

⁸ Yizhuo Ma, Shuang Liang, Rongzheng Wang, Jiakai, Qizhi Chen, Muquan Li, and Ke Qin, “Mitigating KG Quality Issues: A Robust Multi-Hop GraphRAG Retrieval Framework,” arXiv:2603.14828, March 16, 2026. The paper examines retrieval drift caused by noise in imperfect knowledge graphs and retrieval hallucinations caused by incomplete information. https://arxiv.org/abs/2603.14828

⁹ “Limitations of GraphRAG,” Unalarming, 2026. Discusses how single-pass “retrieve once, answer once” systems can miss multi-hop reasoning chains even when the graph contains all required facts, while leaving token cost and latency problems unresolved at scale. https://unalarming.com/limitations-of-graphrag

¹⁰ Cube, “Semantic Layer for AI Agents,” June 3, 2026. Describes the semantic layer as a governed definition of metrics, dimensions, joins, and access policies that allows AI agents to query certified business data rather than repeatedly deriving logic from raw warehouse tables. https://cube.dev/articles/semantic-layer-for-ai-agents-2026

¹¹ Tage Kene-Okafor, “Mem0 Raises $24M from YC, Peak XV and Basis Set to Build the Memory Layer for AI Apps,” TechCrunch, October 28, 2025. The financing consisted of $3.9 million in seed funding and a $20 million Series A. The article reports approximately 13 million package downloads, API calls increasing from 35 million to 186 million within one year, and founder Taranjeet Singh’s observation that major labs have an incentive to keep memory proprietary as a competitive moat. https://techcrunch.com/2025/10/28/mem0-raises-24m-from-yc-peak-xv-and-basis-set-to-build-the-memory-layer-for-ai-apps/