This is part one of the two-part series, "AI for Allocators: Field Notes from a CTO". You can find part two here.
Tetrix is the intelligence platform for capital allocators in private markets. We help institutional investors monitor and collect fund documents from disparate sources, convert this unstructured data into a structured, normalized and queryable format, and provide accompanying insights and analytics. To accomplish this, we run AI in production against millions of financial data points every day, and use it heavily inside our own walls to boost efficiency. Over the past few months, we've seen an uptick in CIOs/COOs of large allocators reaching out to have honest conversations about all things AI. After two years of shipping, breaking things, and rebuilding, we've developed strong views – strong enough that I'm writing them up as a short series.
Here is the argument in one sentence: AI is transformative when applied correctly, but the gap between what's being marketed and what's actually happening is wider than the discourse admits. Two problems in particular deserve more scrutiny than they're getting: how much work it actually takes to extract value from AI today, and how durable the current paradigm will be tomorrow. This piece walks through both, and what to do about them. The next piece walks through why I believe the answer for private markets is vertical AI.
I think we can all agree that AI is here to stay. Capex into AI is exploding, as is the demand for it. On the supply side, Goldman Sachs projects that total capex from hyperscalers – the giant cloud providers like AWS, Microsoft Azure, and Google Cloud that build and run the data centers behind most AI workloads – will reach roughly $1.15 trillion from 2025 to 2027. [1] This is the largest coordinated capital deployment in tech history. On the demand side, IDC's Worldwide AI and Generative AI Spending Guide projects global spending on AI will more than double to reach $632 billion by 2028, a 29% compound annual growth rate.[2] So clearly, there is a very real AI story out there, but it has become a story that crowds out everything else, including the parts that anyone thinking seriously about AI needs to sit with rather than wave away:
Problem #1: The "lackluster / missing ROI" problem
I am not usually one to quote McKinsey, but their 2025 State of AI survey is arguably the most authoritative dataset on enterprise AI adoption. The headlines look great: 88% of organizations now use AI in at least one business function.[3] But the details tell a different story. Only 6% qualify as "AI high performers" – those achieving more than 5% EBIT impact attributable to AI.[3] S&P Global's 2025 Voice of the Enterprise survey of 1,006 IT and business professionals found the gap getting worse, not better: the share of companies abandoning most of their AI initiatives jumped from 17% in 2024 to 42% in 2025, with the average organization scrapping 46% of AI proofs-of-concept before they reach production.[4] And according to PwC's 29th Global CEO Survey, 56% of 4,454 CEOs across 95 countries reported neither increased revenue nor reduced costs from AI over the past 12 months.[5] What I take from these statistics is that there is a wide gap between engineering/deployment reality and commercial storytelling, so let's unpack why that is:
- AI requires system redesign – you can't just buy your team enterprise licenses to copilots and expect magic to happen. For AI to be properly leveraged, institutional processes and workflows need to be fundamentally redesigned. We learned this firsthand. When we shipped our first AI-assisted due diligence module, we bolted on AI to summarize the key points across a set of diligence documents. It worked, but the productivity gain capped out fast. Why? Because the decisions that actually matter in due diligence depend on context, most teams can't easily pull in today – how this manager's pitch today compares to their pitch three funds ago, what comparable GPs in the same vintage have actually delivered, where overlapping exposures already sit across the existing portfolio – because the tooling to do it with speed doesn't exist in their stack. A summarizer didn’t change that; it made the visible slice faster to read, but didn't expand what was visible. The real unlock came when we stopped asking how to speed up the existing DD process and started asking what DD should look like if we built it from scratch around AI. We rebuilt the workflow so the system pre-built comparison matrices across every fund and asset the team had ever evaluated, flagged inconsistencies, and let users rank opportunities against their own criteria. That redesign forced us to reconstruct the layers underneath it – data linkages, entity resolution, analytics interfaces – and the productivity gain only showed up once all three were in place. This pattern is not unique to us. McKinsey's own data backs this up: high performers are nearly three times more likely than peers to have fundamentally redesigned workflows when deploying AI.[3] This is precisely why enterprise AI has perpetually been just around the corner without quite arriving. AI on its own cannot see how work actually happens inside organizations because organizations are not clean systems; they are dense webs of institutional knowledge, undocumented conversations and judgment calls, layered on top of distributed tools that don’t talk to each other and permissioning systems that silo them further. AI can only operate on what it can see, and almost none of that is visible to it. The result, when copilots are rolled out, is predictable: every individual potentially gets a little faster, but when this effort is not coordinated at the system level, the collective output barely moves. Brynjolfsson, Rock, and Syverson's work on general-purpose technologies captures why this dynamic is universal.[6] Firms adopting them typically experience a productivity J-curve, where output dips before it rises, because the real gains come not from the tool but from the slow, costly work of redesigning processes around it. Those who do that work will ride a steeper curve up. Those who don't never climb out of the dip.
- PoC ≠ production grade system – the reason 95% of corporate gen AI pilots fail to deliver measurable P&L impact, per MIT's NANDA initiative, is because people fundamentally underestimate the delta between a proof-of-concept and a production-grade system.[7] As an engineer, I can tell you that gap is not trivial. For instance, you can't just prompt or vibe code your way into a fully secure, scalable and audit-ready portfolio management system. Even if you could, software is not a one-time artifact you generate and walk away from. It is a living system that must be maintained and refined continuously as business needs evolve. This is why the "low code / no code" platform craze in the mid 2010s was not the end of software engineering or SaaS, and why I am also skeptical of LLMs spelling the end of the field as we know it. So yes, LLMs are getting exceptional at producing slick first drafts, but it would be a mistake to think these are finished systems that you can rely on in production, and this is precisely why most AI projects die before they ever produce ROI.
Problem #2: The "evergreen illusion" problem
The AI story of today largely assumes a smooth upward curve: frontier labs maintain dominant positions, models keep getting dramatically better at predictable cost, and the infrastructure to support all this scales as needed. Each of these assumptions is showing visible cracks, and capital is being deployed as if none of them are.
- The moat is thinner than valuations suggest – Consider what the current valuations of the frontier labs actually require you to believe. Every piece of AI capex deployed in the last few years has been underwritten, implicitly or explicitly, by one belief: AI labs have moats deep enough and durable enough to compound for a decade. It is a clean thesis. It is also a thesis that the data has been steadily dismantling. The clearest evidence is in pricing. Per Stanford's AI Index, the cost of GPT-3.5-class inference collapsed roughly 280x in two years – from $20 per million tokens in November 2022 to $0.07 by October 2024.[8] This is not because the AI labs got more generous; it is because the floor keeps dropping out from under them as competitors, and more interestingly, open-weight alternatives force re-pricing. Epoch AI now measures the average lag between open-weight models and the state-of-the-art at roughly three months, down from 5–22 months a year earlier.[9] The DeepSeek release in January 2025 was the moment the market briefly understood what this dynamic meant – a Chinese lab operating under US export controls shipped an open-weight reasoning model that rivaled OpenAI's o1 across math, coding, and reasoning benchmarks, reportedly trained for a fraction of o1's API rate.[10] Then the market chose to forget, even though the delta compression between open and closed source models continues, and the switching costs keep falling alongside it since most AI tech stacks are built to be model agnostic. At the same time, the cost of staying on the frontier keeps climbing. Each generation costs more to train, holds frontier status for less time, and recovers less of its training cost before being commoditized. On top of that, for some workloads, running the models has become more expensive than paying a person to do the same task – as one Nvidia executive recently put it, "the cost of compute is far beyond the costs of the employees."[11]. This further erodes the economic case for adopting top-tier models. Despite all of this, the dominant narrative from frontier labs remains relentlessly bullish on their own durability. That's worth pausing on. Most frontier labs have raised stratospheric levels of capital without a clear-cut path to justifying those valuations. The way then to defend those rounds is by controlling the narrative, whether or not the narrative is grounded in current reality. Reuters' Breakingviews recently coined a term I find comically ironic – "AI revenue hallucination" – for the gap between annualized run-rate claims and actual booked revenue. [12] And this tension is beginning to show even from inside the labs themselves: OpenAI's CFO Sarah Friar reportedly told colleagues she is worried the company may not be able to fund future computing contracts if revenue does not grow fast enough [13]. When the CFO of a category-defining lab is privately worried about the math, the math is worth examining.
- The technical trajectory is less certain than it looks – By late 2024, the original scaling recipe of bigger models, more data, more compute was clearly hitting a wall. The industry pivoted to a new approach called reinforcement learning with verifiable rewards, or RLVR, which powers the "reasoning models" you've heard about: OpenAI's o3, DeepSeek-R1, Gemini Deep Think. The pitch is that RLVR is the new scaling frontier. The emerging evidence is more complicated. For instance, a paper out of Tsinghua University showed that RLVR mostly makes models better at sampling answers their underlying base models could already produce - useful, but not a new scaling law. [14] RLVR also has a deeper limitation: it works where you can mechanically check the answer (math, code, formal logic) and breaks down everywhere else, including most of the nuanced judgment that real business work requires. Meanwhile, the broader debate over whether LLMs are even the right architecture has reopened. Yann LeCun, one of the three "godfathers" of modern AI, left Meta in November 2025 to launch AMI Labs on the bet that LLMs are a dead end and that "world models" trained on sensory data, not text, are the path forward. He raised $1.03 billion at a $3.50 billion pre-money valuation, the third-largest seed round in history, betting against the current paradigm.[15] He may be wrong. He may be right. The point is that the smartest people in the field no longer agree on where this is going, and the truth is that no one really knows for sure. That being said, progress in AI will certainly continue. With so many capable people pushing it forward, backed by ample resources, the opposite view is hard to defend. However, the evergreen illusion assumes a smooth line from here to AGI (assuming we can even get there). The actual line is going to bend, probably more than once, and a lot of today's capital is being deployed as if it won't.
- Scaling atoms is harder than scaling bits – The conversation around AI is often a software conversation, but each year, the binding constraint migrates further into the physical world. Whether it is chip shortages, data center availability, electricity and power grid concerns, the bottleneck is no longer something that software can solve. For example, to manufacture advanced chips, EUV lithography machines are needed, and ASML is the only company on earth that manufactures them. The company plans to ship over 60 EUV systems in 2026, but that is not commensurate with the total commitments being made by all the major players. [16] Power is another big problem. Focusing on the US alone, of the roughly 12 GW of data center capacity announced for delivery in 2026, only about 5 GW is under active construction. In other words, roughly 7 GW has been canceled, delayed, or stalled due to power generation and interconnect constraints.[17] In fact, grid interconnection queues are stretching past five years on the median, with some sectors facing waits of a decade or more.[18] Lead times for high-voltage transformers are also running two to four years.[19] Not to mention that about 70% of US transmission lines and transformers were built between the 1950s and 1970s and are now approaching end of life.[20] The political environment has hardened in parallel too: communities, especially in the US aren't fans of AI and are actively pushing back on data center projects. None of these constraints are unsolvable. But all of them move on physical-world timescales - years to decades, not quarters - and the gap between that clock and the one capital is operating on is the part of the evergreen illusion almost nobody is pricing and will be the most expensive to learn the hard way.
So where does this leave a serious allocator?
To tie a bow to this first piece, I bring up the "lackluster / missing ROI" and "evergreen illusion" problem not to cast doubt on AI itself – we are obviously bullish on and advocates of AI or else we would not be building an AI-native company – but rather to encourage more criticality and thoughtfulness when thinking through the field given all the noise and posturing. For capital allocators specifically, being thoughtful about AI is non-negotiable. Private markets are now estimated to be approaching $20 trillion globally.[21] The volume of unstructured data – fund documents, performance reports, capital calls, side letters – scales linearly with that growth. Your team does not. Without AI-assisted workflows, operational costs become untenable at the scale the industry is heading toward. Even if you discount the hype by 80%, the cost of not having an AI strategy or adopting AI tools is high and rising. The bubble can be real, and the underlying technology can still be transformative for your operations. Both things are true at once, and the hard part – the part that separates sophisticated allocators from the rest – is no longer the technology question, it's the vendor question: evaluating what you're actually being sold.
In the next piece, I'll cover why I believe vertical AI is the right answer for capital allocators in private markets and why horizontal lab offerings won't displace it. Read part 2 here.
Endnotes
[1] "AI boom: Big Tech capital expenditures now seen topping $1 trillion in 2027," CNBC, April 30, 2026. https://www.cnbc.com/2026/04/30/ai-boom-big-tech-capital-expenditures-now-seen-topping-1-trillion-in-2027-.html
[2] "Worldwide Spending on Artificial Intelligence Forecast to Reach $632 Billion in 2028, According to a New IDC Spending Guide," International Data Corporation, August 19, 2024. https://www.businesswire.com/news/home/20240819177906/en/
[3] "The State of AI in 2025: Agents, Innovation, and Transformation," McKinsey & Co., November 2025. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
[4] "Generative AI shows rapid growth but yields mixed results," S&P Global Market Intelligence, Oct. 27, 2025. https://www.spglobal.com/market-intelligence/en/news-insights/research/2025/10/generative-ai-shows-rapid-growth-but-yields-mixed-results
[5] "29th Annual Global CEO Survey: Leading Through Uncertainty in the Age of AI," PwC, Jan. 19, 2026. https://www.pwc.com/gx/en/issues/c-suite-insights/ceo-survey.html
[6] Erik Brynjolfsson, Daniel Rock and Chad Syverson, "The Productivity J-Curve: How Intangibles Complement General Purpose Technologies," American Economic Journal: Macroeconomics, January 2021.
[7] "MIT report: 95% of generative AI pilots at companies are failing," Fortune, Aug. 18, 2025. https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
[8] "The 2025 AI Index Report," Stanford HAI, April 2025. https://hai.stanford.edu/ai-index/2025-ai-index-report
[9] "Open-weight models lag state-of-the-art by around 3 months on average," Epoch AI, Oct. 30, 2025. https://epoch.ai/data-insights/open-weights-vs-closed-weights-models
[10] Ege Erdil, "What went into training DeepSeek-R1?" Epoch AI, Jan. 31, 2025. https://epoch.ai/gradient-updates/what-went-into-training-deepseek-r1
[11] "'The cost of compute is far beyond the costs of the employees': Nvidia exec says right now AI is more expensive than paying human workers," Fortune, April 28, 2026. https://fortune.com/2026/04/28/nvidia-executive-cost-of-ai-is-greater-than-cost-of-employees/
[12] Karen Kwok, "Anthropic Gives Lesson in AI Revenue Hallucination," Reuters Breakingviews, March 2026. https://www.breakingviews.com/columns/considered-view/anthropic-gives-lesson-ai-revenue-hallucination-2026-03-10/
[13] Berber Jin and Tom Dotan, "OpenAI Misses Key Revenue, User Targets in High-Stakes Sprint Toward IPO," Wall Street Journal, April 27, 2026. https://www.wsj.com/tech/ai/openai-misses-key-revenue-user-targets-in-high-stakes-sprint-toward-ipo-94a95273
[14] Yang Yue and others, "Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?" arXiv:2504.13837, April 2025. https://arxiv.org/abs/2504.13837
[15] Anna Heim, "Yann LeCun's AMI Labs raises $1.03B to build world models," TechCrunch, March 10, 2026. https://techcrunch.com/2026/03/09/yann-lecuns-ami-labs-raises-1-03-billion-to-build-world-models/
[16] "Q1 2026 Financial Results," ASML Holding NV, April 15, 2026. https://www.asml.com/en/news/press-releases/2026/q1-2026-financial-results
[17] "Data Center Outlook," Sightline Climate, February 2026. https://www.sightlineclimate.com/research/data-center-outlook
[18] "Queued Up: 2025 Edition – Characteristics of Power Plants Seeking Transmission Interconnection As of the End of 2024," Lawrence Berkeley National Laboratory, December 2025. https://emp.lbl.gov/publications/queued-2025-edition-characteristics
[19] "Power transformers and distribution transformers will face supply deficits of 30% and 10% in 2025," Wood Mackenzie, Aug. 14, 2025. https://www.woodmac.com/press-releases/power-transformers-and-distribution-transformers-will-face-supply-deficits-of-30-and-10-in-2025/
[20] "2025 Infrastructure Report Card: Energy," American Society of Civil Engineers, 2025. https://infrastructurereportcard.org/cat-item/energy-infrastructure/
[21] "Alternative Investments Outlook 2026," J.P. Morgan Asset Management, December 2025. https://am.jpmorgan.com/us/en/asset-management/institutional/insights/portfolio-insights/alternatives/alternatives-outlook/
