The enterprise AI ROI gap persists despite increased agentic spending

confidence 0.5 · horizon 2027-09-04 · status open

Falsifiers

• TODO: what specific, observable outcome would prove this wrong?

Evidence

• +1 Crunchbase: seven billionaire-founded/-led companies took over half of 2026's US venture capital — New granular data on extreme capital concentration (seven billionaire-founded companies, over half of 2026 US VC) extends the already-tracked pattern of capital pouring into AI well ahead of resolved enterprise ROI.
• +2 Anthropic's compute margins reportedly worsening rather than improving — New, specific reporting that a frontier lab's compute margins are worsening rather than improving extends the already-dense capex/cashflow-strain evidence chain specifically for Anthropic (prior cloud-partnership and credit-rating items), though the available text is a brief teaser without full figures.
• +2 OpenAI closes $11B raise at $1.25T valuation, well short of its $30B target, as investors push Anthropic toward a lower IPO price — A concretely smaller-than-sought OpenAI raise plus reported downward pressure on Anthropic's IPO valuation extends the already-tracked pattern of capital growing more selective toward frontier-lab valuations, though relayed via a single critical-leaning outlet's digest of WSJ/Bloomberg reporting rather than the primary articles.
• +1 Crunchbase: global VC funding hit $59B in October, driven by DeepSeek's record $15.9B round; Moonshot AI reportedly reaches $50B valuation and files confidentially for a Hong Kong IPO — Continued heavy concentration of global VC funding into a handful of AI megarounds extends the already-tracked pattern of capital pouring into AI companies well ahead of resolved enterprise ROI questions.
• +3 US midterm elections deliver nationwide electoral rejection of AI data centers — A nationwide electoral rejection of AI data centers, including a governor's defeat tied directly to his pro-data-center record, is the strongest capex-constraining political signal yet in the buildout-outrunning-capacity pattern this thesis already tracks.
• +1 AI data center builders turn to decommissioned jet engines for power amid grid delays — Operators resorting to decommissioned jet turbines for power is a vivid new instance of the already-tracked pattern that grid/power delivery, not capital, is constraining AI buildout.
• -1 OpenAI publishes Heimstaden case study on ChatGPT Enterprise and Codex for property management — Another named-enterprise customer success story is mild counter-evidence to the persistent-ROI-gap thesis, though it is OpenAI's own unaudited account with no disclosed figures, matching the already-tracked pattern of similar vendor case studies.
• -1 AI inference startup Fal doubles valuation to $16B on demand surge — A demand-surge-driven valuation doubling for an inference infrastructure startup is mild counter-evidence that some AI infrastructure layers have real, paying usage, similar to the already-tracked ElevenLabs and Modal items.
• +2 xAI throttles Grok Bot usage limits, citing 'unsustainably high' costs — A named lab leader's own admission that a fast-growing agentic product's costs are 'unsustainably high' enough to force throttling is direct, named-source evidence that agentic-AI economics remain strained even for high-adoption products, extending the already-dense cost/ROI evidence chain.
• +1 Crunchbase: Q3 VC funding hits record dollars but fewer deals, concentrated in AI megarounds — Fresh Q3 figures showing a record $680B in global VC funding concentrated into a small number of AI megarounds, while earlier-stage deal counts fall, extends the already-tracked pattern of capital pouring into AI companies well ahead of resolved enterprise ROI questions.
• +1 OpenAI launches utility partnership program amid talk of rebranding data centers — OpenAI's own utility-partnership program and PR rebranding attempt is further evidence that AI infrastructure buildout is generating enough community/political friction to need dedicated mitigation programs, extending the already-tracked pattern (Stargate force majeure, Nscale Chestnut Grove dispute, Texas permit shutdown).
• -1 Starbucks, Wingstop, and Chipotle expand AI investments for customer insights and kitchen efficiency — Continued, if vague and unquantified, real-world enterprise AI investment across several named chains is mild counter-evidence to the persistent-ROI-gap thesis, similar in weight to the already-tracked Chewy and BofA items.
• +2 OpenAI reportedly lowers funding target for its new round — A reported pullback in OpenAI's own fundraising ambition is a demand-side crack in the AI capital supercycle, though the report lacks specific figures.
• +1 Judge allows NAACP's Clean Air Act lawsuit against xAI data center to proceed — A federal judge allowing a Clean Air Act lawsuit against xAI's Memphis data center to proceed extends the already-tracked pattern of AI infrastructure buildout facing real-world regulatory and community friction, alongside Nscale's Chestnut Grove dispute and the Texas data center permit-portal shutdown.
• +2 Proton cancels AI "re-interviews" after 404 Media exposes culture of fear and revenue-driven AI mandate — An internal document disclosing Proton is 'by all metrics that matter...an underperforming company' missing its ARR target, used as explicit justification for a coercive AI mandate employees say compromised security review, is direct evidence of revenue-driven, performative AI adoption extending the pattern already tracked via Visier's survey and the Claude-Code-mandate and Accenture-spend-cap items.
• +1 DoorDash says it has received only "several hundred" orders from AI agents — A named platform's own disclosed figure — only several hundred AI-agent orders to date — is a concrete, if narrow, data point that agentic-commerce adoption lags the hype well behind announced capability.
• +2 Nvidia weighs cosigning financing for Nscale's West Virginia AI data center — A chip vendor weighing cosigning a customer's data-center debt is evidence that capex-outrunning-returns strain is reaching credit markets, extending the already-tracked capex/cashflow chain.
• +1 Intel stock crash erases $12B from the US government's equity stake — AI-driven memory and component demand squeezing Intel's PC business and a government equity stake extends the already-tracked component-cost-pressure pattern (Gartner's 2027 memory-cost warning) with a quantified loss.
• +2 Anthropic expected to receive a high 'junk' credit rating from agencies after its IPO, per FT report — A forecast sub-investment-grade rating raising Anthropic's cost of debt extends the already-dense pattern of frontier-lab financial strain disclosed via its S-1 and reported cloud-partnership difficulty.
• +2 Alphabet suffers its worst day ever as a public company amid AI capex worries and an admission that AI Mode makes less money than search — A hyperscaler's own CBO publicly confirming its flagship AI product earns less revenue per use than the product it's replacing, alongside a swing to negative free cash flow, extends the already-dense capex/cashflow-strain evidence chain to Google specifically.
• +1 AWS CEO says the cloud unit will become a $1 trillion business, but only after further heavy capex — A third major cloud provider publicly committing to continued heavy AI capex ahead of proven returns, in the same week as Alphabet's capex-linked selloff, extends the already-tracked pattern of capital outrunning demonstrated ROI.
• +1 Major banks reportedly back Anthropic's compute buildout — Commercial banks reportedly joining the roster of financiers backing Anthropic's compute buildout would extend the already-tracked pattern of capital pouring into AI infrastructure ahead of proven ROI, though the only available source is a thin single-outlet headline.
• -1 Microsoft discloses Azure's dollar revenue for the first time as both cloud giants report AI-driven growth — A first-ever dollar-revenue disclosure showing real, substantial cloud revenue growth is vendor-side counter-evidence to the pure capex-outrunning-revenue reading, similar to the already-tracked Anthropic Q3 revenue disclosures, though it is infrastructure revenue rather than proof of enterprise-customer ROI.
• -1 Bank of America and Capital One say smaller AI models are helping manage costs — Named bank executives describing successful cost control through deliberate smaller-model tiering is mild counter-evidence to the persistent-ROI-gap thesis, similar to already-tracked efficiency claims from LangChain, Cohere, and AI21, though no specific savings figures are given.
• +1 Gartner: non-technical executives place less importance on AI shaping business strategy than CEOs and CIOs — A named analyst firm's finding that non-technical executives undervalue AI's strategic role relative to CEOs/CIOs is mild operational-lane evidence of the cross-functional misalignment consistent with the persistent-ROI-gap thesis, though the available text gives no magnitude or survey methodology.
• +1 Anthropic officially files S-1 with the SEC for proposed IPO — Anthropic's own S-1 formally discloses mounting losses alongside soaring revenue, converting previously leaked prospectus figures into an official regulatory filing.
• +1 SpaceX locks in $40B GPU financing deal with Apollo — A further large AI-infrastructure financing commitment ahead of proven enterprise ROI extends the already-tracked pattern of capex continuing to outrun demonstrated returns.
• +1 Proton tells staff it will 're-interview' them about AI usage, with continued employment potentially at stake — A documented internal policy making continued employment partly contingent on demonstrated AI usage, rather than any measured output or productivity outcome, extends the already-tracked pattern of performative/forced AI adoption (the earlier Claude-Code-mandate anecdote and Visier's performative-AI-use survey) consistent with the persistent-ROI-gap thesis.
• -2 Anthropic's S-1 reveals 14x H1 revenue growth, a Q2 operating profit, and a $120.5B Nvidia compute commitment — Anthropic's own disclosed swing to a Q2 operating profit ($974M) and an 85% H1 gross margin is direct, filing-sourced evidence that at least one frontier lab's own unit economics are improving, cutting against the pure capex-outrunning-revenue reading, though the roughly $80B H1 net loss (mostly non-cash SAFE/derivative revaluation) still shows reported losses are large in GAAP terms.
• +2 Apollo economist says AI has delivered "zero financial payoff" to most companies outside tech — A named investment firm's economist asserting AI has produced zero financial payoff for corporate America outside tech is a credible, if thinly-sourced-in-this-blurb, data point directly supporting the persistent-ROI-gap thesis.
• +1 AI lab Flapping Airplanes raises $1B Series A for cheaper-to-train, cheaper-to-run models — Another large, concentrated capital bet on an unproven, pre-product AI lab extends the already-tracked pattern of capital pouring into AI companies well ahead of demonstrated enterprise ROI.
• +1 Anthropic reported to delay planned IPO as OpenAI's Altman rules out a 2026 listing — New granular loss figures ($42B net loss, $7.33B compute spend against $4.6B revenue) extend the already-tracked pattern of frontier-lab economics not yet working, though relayed via a single opinionated outlet.
• +1 Texas shuts down automated data-center permit portal, suspends 17 permits including OpenAI and Amazon projects — Another instance of AI infrastructure buildout running into real-world regulatory/process friction that outran governance capacity, extending the already-tracked pattern (Oracle Stargate, Crusoe, Nscale Chestnut Grove).
• +1 "Black October": AI-driven attack framework breaches ConnectWise and ripples into a wider vulnerability-disclosure wave — CIO Dive's reporting on tech executives citing overreaching autonomous agents and unclear C-suite accountability extends the already-dense operational-accountability-gap evidence chain.
• +2 OpenAI reportedly working on new $30B raise at $1.25T valuation ahead of IPO — A record-setting $1.25T private valuation ahead of IPO extends the already-tracked pattern of capital pouring into AI companies well ahead of resolved ROI and disclosed-loss questions (Anthropic's prospectus losses, OpenAI's own revenue-shortfall reporting).
• -1 AI21 says it cut GitLab Duo Code Review's AI cost by 98% — A quantified 98% cost-reduction claim is a mild efficiency counter-data-point, though self-reported and unaudited like similar prior claims from LangChain and Cohere.
• -1 OpenAI launches Decisions API, entering the cheap 'decision model' routing category — OpenAI's own low-cost decision-model entrant is mild efficiency counter-evidence to the ROI-gap thesis, similar to the already-tracked LangChain/Jev and Cohere cost-cutting claims.
• -1 Modal crosses $1B ARR, raises $750M Series D at $16B valuation on agentic-sandbox demand — A profitable, organically-growing AI infrastructure vendor crossing $1B ARR on real agentic-sandbox demand is mild counter-evidence to the persistent-ROI-gap thesis, similar to the already-tracked ElevenLabs valuation item.
• +2 Linux Foundation survey: AI-generated 'slop' now a quarter of open-source pull requests — A named Linux Foundation survey quantifying AI-generated 'slop' pull requests (about a quarter of incoming PRs) and maintainer strain severe enough that the curl project changed its disclosure policy is direct operational-lane evidence that AI coding assistance is creating real costs alongside its claimed time savings.
• +2 Survey: over half of managers say AI made evaluating employee work harder — A survey finding that over half of managers now find it harder to evaluate employee work after adopting AI is direct operational-lane evidence consistent with the persistent-ROI-gap thesis.
• +1 Gartner warns AI demand will push memory costs higher into 2027 — Gartner's warning that AI-driven memory costs will rise into 2027 is mild forward-looking cost-pressure evidence consistent with the persistent-ROI-gap thesis.
• +1 Historic chestnut grove sits in footprint of Nscale data center campus serving Anthropic, Nvidia, Microsoft — Community and environmental resistance to a data center campus serving Anthropic, Nvidia, and Microsoft extends the already-tracked pattern of AI infrastructure buildout facing real-world delivery friction.
• +1 Meta plans a new class of stock to help fund its AI buildout — A purpose-built new stock class to fund AI buildout is a financing-structure response consistent with capex continuing to outrun demonstrated returns.
• +2 Splunk: more than half of businesses are failing to control AI costs amid agentic spending — A quantified majority-level AI cost-control failure extends the already-dense ROI-gap evidence base, though it is a single vendor's self-interested report.
• +2 Anthropic reportedly struggling to secure enough cloud compute ahead of planned November IPO — Anthropic reportedly struggling to secure enough compute from cloud providers that increasingly see it as a competitor, even as infrastructure spend balloons ahead of a planned November IPO, extends the already-tracked capex/cashflow-strain evidence chain.
• -1 SpaceX reports Q3 profitability, crediting its AI business — A reported AI-driven profitability milestone (net income of $4.4B) is mild counter-evidence to the persistent-ROI-gap thesis, though the reporting is thin on detail and it is unclear how directly this reflects xAI's own model economics versus SpaceX's broader business.
• +1 Gartner: leaders underestimate the hidden cost of delegating work to AI — A named analyst firm flagging hidden costs of AI delegation is mild operational-lane evidence consistent with the persistent-ROI-gap thesis, though the available coverage is a single-sentence teaser with no specific findings.
• +2 Gartner: CIOs report misaligned AI cost/trust and almost no plans to retool for an agentic workforce — A named analyst firm's conference findings plus a survey showing only 4% of tech leaders have agentic-workforce retooling plans is credible operational-lane evidence consistent with the persistent-ROI-gap thesis.
• +1 Nscale and Anthropic reportedly in talks for a $45B compute deal — Another large reported compute commitment ahead of proven ROI on such spend extends the capex-outrunning-returns pattern, though it is unconfirmed 'in talks' reporting.
• +1 Disney agrees to hire majority of Character.AI's workforce in bankruptcy stalking-horse bid — Disney absorbing most of Character.AI's workforce via bankruptcy extends the already-tracked Character.AI economics evidence with a concrete dissolution outcome rather than a turnaround.
• +1 Asia AI startup funding holds near record in Q3 2026, led by Moonshot AI's $3.5B and Kling AI's ~$3B rounds — Continued large-scale AI capital formation in Asia, concentrated in a handful of jumbo rounds, extends the already-tracked pattern of capital pouring into AI companies regardless of demonstrated enterprise ROI.
• +2 Replit stops using Anthropic models, CEO cites cost among reasons — A named CEO citing cost as a reason for dropping a frontier model provider is concrete, named-source evidence of cost pressure shaping real agentic-coding purchasing decisions.
• +2 CIO Dive: agentic AI success gap persists between executives and workers — A named-outlet finding that executives report more agentic-AI success than the employees actually using the tools is direct operational-lane evidence that perceived gains are outrunning ground-level adoption.
• -1 Jev becomes Vercel's fastest-adopted AI model ever, per AI Gateway usage data — Rapid, gateway-measured developer adoption of a cost-efficient decision model extends the already-tracked mild counter-evidence (LangChain's Jev cost-cutting claim) that some agentic-AI efficiency gains are translating into real uptake, though Vercel discloses no actual usage numbers.
• -1 OpenAI publishes Itaú generative AI customer story — A named enterprise customer story is mild counter-evidence to the persistent-ROI-gap thesis, though it discloses no quantified figures and is OpenAI's own customer account.
• +1 Revenera study: software vendors and CIOs still lack visibility into whether AI features are paying off — A named vendor study finding both software providers and CIOs lack visibility into whether AI features are paying off extends the already-dense pattern of measurement-gap evidence (Pluralsight, TrustedTech, Accenture, Gartner) supporting the persistent-ROI-gap thesis.
• +2 Wharton study: dedicated AI officers and platforms link to better ROI, but clear ownership matters more than titles — A named-study data point isolating ownership/governance, not titles or platforms, as the deciding ROI variable directly extends the already-dense operational-lane accountability/ownership evidence chain (FICO's platform-sprawl finding, Collibra's accountability survey, Gartner's governance-model finding).
• +1 Modal raises $750M Series D at $15.75B valuation, crosses $1B ARR — Another large, quantified AI-infrastructure funding round continues the pattern of capital pouring into the agent/compute layer, alongside already-tracked Crusoe/Nscale/Mistral/Lambda rounds, though as an infra vendor's own revenue growth it is supply-side demand evidence rather than proof of enterprise ROI.
• -1 Cohere Labs reports teams of coding agents can reach frontier coding quality at 30% of the cost — A vendor-reported method for cutting multi-agent coding costs to roughly 30% of typical spend is a mild efficiency counter-data-point to the persistent-ROI-gap thesis, similar to the already-tracked LangChain decision-models claim, though self-reported and unaudited.
• -2 Anthropic tells investors Q3 revenue hit $16.8B, up ~50% quarter over quarter, ahead of IPO — A disclosed $1.70-per-$1-compute revenue ratio and 50% QoQ growth is vendor-side counter-evidence that frontier-lab economics aren't working, cutting against the capex-outrunning-revenue pattern otherwise dominant this cycle, though it is lab-side economics rather than enterprise-customer ROI.
• +1 OpenAI and Anthropic ARR figures draw scrutiny over undefined, inconsistent accounting methodology — Zitron's relay of FT reporting that OpenAI's and Anthropic's ARR figures rest on undefined, inconsistent accounting conventions extends the already-tracked pattern of AI-industry financial opacity, though single-sourced via a critical-leaning outlet relaying FT reporting rather than the primary FT article in hand.
• +2 Character.ai bankruptcy filings show AI compute costs persistently outpaced revenue — Character.ai's bankruptcy filings are a rare, legally-documented instance of an AI company's compute costs persistently exceeding revenue even after aggressive cost-cutting, extending the already-tracked vendor-side economics-not-working pattern (alongside Anthropic's and OpenAI's own disclosed losses/shortfalls).
• +1 Amazon ends NDA secrecy on AI data center deals, following Microsoft — Community backlash forcing a second hyperscaler to drop deal secrecy extends the already-tracked pattern of AI infrastructure buildout facing real-world friction.
• +1 Zitron: AI data center buildout is increasingly financed through junk-rated debt — Continues the already-tracked capex/cashflow-strain evidence chain with new junk-bond-market figures, though the credit-rating characterization is single-sourced opinion.
• +1 Safe Superintelligence raises $5B round with Nvidia, leading Q3 2026's AI megadeals — Continued multi-billion-dollar frontier-lab funding even as broader AI venture funding cools extends the pattern of capital outrunning demonstrated enterprise ROI.
• +1 Flex spins out $2B-backed AI cloud and power infrastructure firm Axiom Solutions, plans IPO — A further $2B commitment to AI-adjacent cloud/power infrastructure, valued at $37.5B enterprise value ahead of a planned IPO, extends the already-tracked pattern of capital outrunning demonstrated enterprise AI ROI.
• -1 Sophos says OpenAI's Daybreak cuts threat-investigation time 96%, automates 52% of MDR cases — A named enterprise reporting large, specific automation gains is mild counter-evidence, though it is OpenAI's own unaudited customer story.
• -1 TypeSafe AI's Jev hits $7.5B valuation in three weeks as a 'decision model' category forms across OpenAI, Microsoft, Perplexity, and Cloudflare — LangChain's claim that routing yes/no agent steps to a decision model cut median cost per task by 64% is a mild efficiency data point cutting against the persistent-ROI-gap thesis, though it is a single vendor's self-reported routing-cost claim relayed via a curator digest, not independently audited.
• +2 OpenAI's annualized revenue reportedly $20 billion below earlier $70 billion projection — A specific, large downward revenue revision is direct financial evidence that AI-industry economics are not matching the growth narrative.
• +1 Thoughtworks report: no dominant model for enterprise AI governance, CIOs feel accountable for its mistakes — A named-firm finding that enterprises lack a dominant AI governance model, with CIOs feeling personally accountable for AI's mistakes, extends the already-tracked operational-lane accountability-gap pattern.
• +2 CIO Dive: AI platform sprawl is undermining enterprise AI ROI — A named executive (FICO) attributing poor enterprise AI ROI to platform sprawl rather than model capability is a fairly direct, named-source operational-lane data point for the persistent-ROI-gap thesis.
• +1 Survey: executives use AI more than their workers, many admit overstating their own AI knowledge — Executives overstating their own AI fluency while outpacing their workers' AI use is mild operational-lane evidence consistent with the persistent-ROI-gap thesis, since it implies adoption decisions rest on overstated expertise.
• +1 Zitron: rising Treasury yields raise the cost of AI's coming debt wave, citing Anthropic's and Oracle's exposure — Zitron's tracking of rising Treasury yields raising the cost of AI labs' and data centers' coming debt needs extends the already-tracked hyperscaler/lab capex-cashflow-strain evidence chain, though it is single-sourced opinion analysis.
• +1 AI compute startup Lambda raises $4B ahead of planned 2027 IPO — Continued large capital inflows to AI compute infrastructure ahead of Lambda's own planned IPO extends the already-tracked pattern of AI capex continuing to outrun demonstrated enterprise ROI.
• +1 Crunchbase: Q3 2026 set a record for billion-dollar venture rounds, AI still 64% of global VC — A record pace of billion-dollar AI infrastructure rounds (Crusoe, Nscale, Mistral, others each $3B+) extends the already-tracked pattern of AI capex continuing to outrun demonstrated enterprise ROI.
• +1 Bain & Co.: AI coding speed gains are being absorbed by code-review bottlenecks — A named consultancy partner's finding that the coding-agent bottleneck has shifted from generation speed to review/trust capacity is operational-lane evidence consistent with the persistent-ROI-gap thesis, since it implies raw throughput gains don't cleanly convert to realized value.
• +1 Anthropic commits $100 million to train 10,000 engineers on enterprise AI — Anthropic's own $100M investment to close the enterprise AI talent gap implicitly acknowledges the skills-gap bottleneck already tracked via Pluralsight and TrustedTech data, though it is a vendor-interested action rather than independent evidence.
• +2 Goldman Sachs: AI investment to reach 1.9% of US GDP in 2026, almost entirely data-center buildout — A named, credible estimate (Goldman Sachs via WSJ) that AI's GDP contribution is almost entirely data-center capex rather than AI revenue extends the already-tracked capex/cashflow-strain evidence chain (hyperscaler capex, Stargate, Crusoe), though the capex-vs-revenue framing is Zitron's own analysis of BEA categories rather than a primary Goldman report in hand.
• -2 Airbnb CTO details AI-native transformation: 60% AI-authored code, ~45-50% of support tickets resolved by AI — Airbnb's named, quantified figures (60% AI-authored code, 1.6x PR throughput, ~45-50% of support tickets resolved by AI) paired with deliberate process redesign is credible counter-evidence to the persistent-ROI-gap thesis, though the figures are self-reported by the company rather than independently audited.
• -1 KPMG: finance teams get the most AI value from judgment-heavy tasks — KPMG data showing AI performs well on judgment-heavy finance tasks is mild counter-evidence to the persistent-ROI-gap thesis, though it's narrow to one function and doesn't speak to overall AI spend ROI.
• +2 Gartner forecasts most vendor forward-deployed-engineer programs abandoned by 2028 — A credible analyst-firm forecast that most vendor FDE programs will be abandoned by 2028 over cost and efficacy problems is direct operational-lane evidence consistent with the persistent-ROI-gap thesis, though it is a forward forecast rather than realized attrition data.
• +1 Nuclear startup funding hits record high on AI power demand as recent nuclear IPOs slump — Sharp post-IPO declines in AI-power-adjacent nuclear startups even as private nuclear funding hits a record extend the already-tracked capex/cashflow-strain and buildout-outrunning-capacity evidence chain, though this is energy-sector data adjacent to AI rather than AI-lab spending directly.
• +2 Pluralsight: skills gaps, not tooling, wasted $250K+ at most businesses in the past year — A quantified figure ($250K+ wasted per business from skills-gap-driven project misalignment) is direct operational-lane evidence consistent with the persistent-ROI-gap thesis, though it rests on a single vendor-commissioned report.
• +2 Accenture caps employee Claude spending at $29/month amid AI token cost overruns — A quantified employee-level Claude spend cap after heavy usage, imposed by a consultancy that sells AI-transformation advice to clients, is direct evidence of the cost side of the persistent ROI tension, extending the already-dense evidence chain on enterprise AI cost/value mismatch.
• -1 Barclays scales Claude deployment across operations — A named enterprise's vendor-reported deployment story is mild counter-evidence to the persistent-ROI-gap thesis, though it discloses no figures and is Anthropic's own customer story.
• -1 ElevenLabs valuation doubles to $22B via employee tender — A doubling to a $22B valuation via a large employee tender for a profitable-product AI company is mild counter-evidence to the persistent-ROI-gap thesis, though ElevenLabs is a narrow vertical-product company rather than a general enterprise AI deployment, limiting how much it generalizes.
• +2 Morgan Stanley: over half of AI GPU servers shipped 2026-2028 may lack power to plug in, as GPU rental rates stay flat — A named-analyst estimate that over half of 2026-2028 GPU shipments may lack usable power extends the already-tracked pattern of AI infrastructure buildout outrunning executable delivery capacity (alongside Oracle Stargate and Crusoe items), with a specific magnitude attached this time, though it arrives via Zitron's relay of FT/Morgan Stanley/Jefferies reporting rather than a primary research note.
• +1 GFT Technologies report: modernization costs and unclear vision are holding back enterprise AI adoption — A vendor-commissioned survey finding that modernization cost and unclear vision, not AI tooling itself, are holding back adoption is operational-lane evidence consistent with the persistent-ROI-gap thesis, though it is a single vendor-sponsored report.
• +2 Anthropic's IPO prospectus discloses tens of billions in annual losses alongside rapid growth and an AI existential-risk warning — Anthropic's own prospectus-disclosed tens-of-billions-per-year losses is direct evidence that frontier-lab economics haven't reached sustainability even at rapid revenue growth, extending the supply-side capex/monetization-gap pattern already tracked via hyperscaler capex and Stargate/Crusoe items, though the figures are TechCrunch's characterization of the filing rather than the primary numbers in hand.
• +1 Blue Cross Blue Shield claims hospital AI tool use added $942M in healthcare costs over two years — An insurer's claim that hospital AI tool use is adding measurable cost rather than delivering the promised efficiency gains is weak, single-sourced, interested-party evidence adjacent to the persistent-ROI-gap thesis, though it measures cost inflation rather than direct productivity ROI.
• +1 Enterprise AI agents remain siloed even as investment in them accelerates, per CIO Dive — A vendor executive's characterization of enterprise agents as disconnected even as investment accelerates is consistent with the persistent-ROI-gap thesis, though it rests on a single vendor's framing rather than independent survey data.
• +1 Anthropic commits $11.6B over seven years to Akamai for CPU cloud infrastructure, with equity stake attached — A large, long-term compute commitment made ahead of proven ROI on such spend adds to the pattern of AI capex continuing to outrun demonstrated returns, alongside already-tracked hyperscaler capex/cashflow evidence.
• +1 Crusoe abandons $1.25B plan to use Boom Supersonic turbines at AI data centers — A named AI infrastructure operator publicly abandoning a large power-sourcing commitment is further supply-side evidence that AI buildout plans are outrunning executable delivery capacity, alongside already-tracked Oracle Stargate and hyperscaler capex evidence.
• +1 AWS research warns CIOs risk wasting resources without a clear organizational path to AI value — AWS's own research warning that CIOs risk wasting resources without a clear organizational path to value is a vendor-sourced but notable acknowledgment consistent with the persistent-ROI-gap thesis.
• +1 Meta doubles down on Muse as its "centerpiece," bringing it to AI glasses and a new wearable "Charm" device — Meta committing to a large hardware and capex push around a month-old, still-unproven agent product is mild evidence of AI investment continuing to outrun demonstrated returns.
• +1 Oracle sends force majeure notice on its New Mexico Stargate data center, signaling delivery slippage — A force majeure notice on a named Stargate site is adjacent supply-side evidence that AI infrastructure buildout may be outrunning executable delivery capacity, alongside already-tracked hyperscaler capex/cashflow strain evidence.
• -1 Bank of America to double its AI budget next year, citing 'clearly identifiable' gains — A large bank's public commitment to double its AI budget while citing 'clearly identifiable' gains is mild counter-evidence to the persistent-ROI-gap thesis, though the framing is hedged by CIO Dive's own note that ROI-timing questions remain, and no specific figures are disclosed.
• +2 Visier study finds employees report 'performative' AI use amid job-loss fears — A named survey finding that enterprise AI use is often performative — driven by pressure and job-loss fear rather than genuine adoption — is direct operational-lane evidence consistent with the persistent-ROI-gap thesis, since activity metrics under this dynamic would overstate real productivity gains.
• +1 Zitron: Bloomberg data shows only ~2 of Microsoft's 12 gigawatts of data center capacity is AI-specific — Bloomberg-reported data that only about 2 of Microsoft's 12 gigawatts of data center capacity is AI-specific is mild supply-side evidence that AI infrastructure buildout claims may be outrunning usable capacity, adjacent to the persistent-ROI-gap thesis, though it rests on a critical-leaning source's interpretation of secondhand reporting.
• +2 McKinsey: most enterprises overrunning AI budgets on coding agents despite lackluster results — McKinsey-sourced reporting that most enterprises are overrunning AI budgets on coding agents despite lackluster results is a fairly direct, named-source data point for the persistent-ROI-gap thesis.
• +1 Anonymous account: company where all specs and code are AI-generated, staff working 12-13 hour days "to press enter" — A single anonymous but vivid anecdote of mandated Claude-Code-only development producing exhaustion without leadership perceiving any productivity bottleneck removed is weak operational-lane evidence consistent with the persistent-ROI-gap thesis, capped low given it is unverified and single-sourced.
• +1 Survey finds skills gaps, not compute, are blocking enterprises from full AI gains — A survey finding that training gaps rather than compute block enterprises from AI gains is operational-lane evidence consistent with the persistent-ROI-gap thesis, though it measures skills readiness rather than direct financial ROI.
• +1 Zitron: Oracle, Google, and Amazon have gone cashflow-negative funding the AI buildout, with Meta close behind — Zitron's tracking of hyperscaler cashflow strain from AI capex is adjacent supply-side evidence that AI monetization isn't matching investment, though it is single-sourced opinion analysis without a primary earnings citation.
• +1 Survey: over half of enterprises still building internal accountability lines for AI outputs; failures traced to data-foundation gaps — A survey finding that most enterprises still lack clear internal accountability lines for AI outputs, with failures traced to data-foundation gaps, is a mild operational-lane data point consistent with the persistent-ROI-gap thesis, though it measures process maturity rather than direct ROI.
• +2 Databricks reports 60% higher coding-agent spend after rolling GPT-6 Astra out to ~3,500 engineers — A named executive's report that total coding-agent spend rose 60% after adopting a per-task-cheaper model is concrete evidence that per-task cost claims don't guarantee lower total spend, though it is social-media-relayed rather than a published case study.
• +1 Steve Yegge shuts down his 'Gas Town' multi-agent coding orchestrator, citing reliability problems — A heavy individual spender abandoning a multi-agent orchestrator after failing to get proportional output beyond the tool itself is mild first-person evidence toward the persistent-ROI-gap thesis.
• +2 Accenture data: only 1 in 5 enterprise AI token-spend dollars tied to measurable business outcomes — Accenture-reported data that only 1 in 5 enterprise token-spend dollars ties to tangible results is a concrete quantified data point supporting the persistent-ROI-gap thesis, though it rests on a single press report of one consultancy's figure without visible methodology.
• -1 Chewy projects $50M in annual cost savings from AI deployments across customer care, pharmacy, and vet services — A named enterprise's quantified $50M annual AI cost-savings projection is mild counter-evidence to the persistent-ROI-gap thesis, though it is a self-reported target rather than audited results.
• +2 AI spend per employee reportedly fell at top firms in August — A reported per-employee AI spend decline at top firms in August, framed by TechCrunch as a possible warning sign rather than pure seasonal softness, is evidence toward the persistent-ROI-gap thesis.

Confidence history

• 2026-09-04 → 0.5 — Initial — set on adoption; owner to adjust.

← All theses