ClarityBeta

Investment & Market Intelligence

The Investor

12 min read · 2,407 words · 4 moves to watch
MONDAY, SEPTEMBER 7, 2026
The Signal

Only 13% of US power capacity queued from 2000 to 2020 was running by end-2025.

Another 549 GW sits under signed interconnection agreements and still isn't operating, which is awkward arithmetic when a data center campus takes roughly eighteen months to build. Which means any mark underwritten on announced megawatts is really a five-year bet on on-site gas, and the sponsors' own filings say as much. The bull case assumes the queue clears in time; Georgia's docket adds a litigation fight the current marks assume away.

In Play

  1. Interconnection Queue Attrition Reprices Data Center Marks

    Berkeley Lab data show that of all US generation capacity entering interconnection queues from 2000 to 2020, only 13% was operating by the end of 2025. A further 549 GW holds a draft or executed interconnection agreement and still is not running, and 2025's completions waited a median of more than five years. For your data center and neocloud marks, announced pipeline megawatts describe a queue position rather than delivered power. Anything on the wrong side of that gap carries implicit gas exposure for its first half-decade.

  2. Valuation Methodology Becomes a Legal Document

    The Justice Department opened a probe into how BlackRock's TCP Capital valued its assets, after the fund cut net asset value 19% in January and a further 5% in May. Separately, KPMG warned a subsidiary of Guggenheim's $367bn asset manager over internal-control deficiencies tied to how it recognized hundreds of millions of dollars of revenue, following a whistleblower complaint, per the Financial Times. Valuation methodology has moved from an audit conversation to a discoverable legal record, and your Q3 letters ship into that.

  3. Inference Margin Is Set in the Checkpoint, Not in Ops

    Anthropic cut cache-read pricing 75%, which it estimates lowers typical workload cost by roughly 25% and more on agentic work. Meta's Muse Spark 1.3 scores 61 on the independent Artificial Analysis index at $1.25 and $4.25 per million tokens, against GPT-6 Astra's $10 and $50 for capability that index cannot distinguish. Yet Claude Fable 5.1's cost per task rose about 20% generation over generation anyway. Per-token price has stopped predicting portfolio gross margin.

  4. Youth Sports Growth Is Price, Not Players

    Family spending on youth sports rose 46% between 2019 and 2024 while one in five parents limited or ended a child's participation, in a market the Aspen Institute sizes at $40B a year. KKR's Varsity Brands, the sector's flagship outcome, went from $1.5B in 2014 to $4.5B in 2024 and paid $82.5M to settle anticompetitive allegations along the way. Growth here is average revenue per family rather than volume, and Senator Chris Murphy is drafting private-equity restrictions into the same window 2020-2022 vintages need liquidity.

  5. The Verification Layer Is Fully Documented and Still Unfunded

    Netflix published its complete production LLM-judge lifecycle: rubric tuning that treats a right answer for the wrong reason as its own error class, a Meta-Judge comparing the judge's rationale to a human rater's, and drift bands calibrated against human inter-rater spread. It shipped no product. ServiceNow's AgentJudgeBench found judge alignment degrades on hard tasks, and Qwen's E-Commerce Bench showed agent profitability across 18 models does not correlate with fraud avoidance or negotiation quality. Enterprises cannot verify agent behaviour, which caps agent revenue expansion regardless of model capability.

Deep Dives

Announced Megawatts Are a Milestone, Not a Delivery Date

The sector's favorite valuation input is a queue position with an 87% historical failure rate, and the workaround — on-site gas — now carries litigation risk that can strand a campus mid-ramp.

Why the gap fills with gas

Clean generation needs five-plus years to energize and a data center campus takes roughly eighteen months to build, which settles the fuel mix before anyone gets to argue about policy preference. Georgia is the clearest published case. Georgia Power's 2025 resource process authorized up to 8.5 GW of new capacity, and the Georgia Public Service Commission said roughly 80% of the December 2025 approvals will serve data centers. The certified portfolio comes to 9,885 MW: 3,692 MW of new combined-cycle gas, 2,762 MW of standalone batteries, 350 MW of solar-paired storage, 260 MW added to an existing solar power purchase agreement, and 2,821 MW of other PPAs. Project and transmission spend runs to about $16.7 billion. The same plan pushed Plant Scherer Unit 3 and Plant Bowen Units 1–2 from a planned 2028 retirement to at least 2035.

Singapore ran the same play in a different jurisdiction, awarding 670 MW of gas to Tuas Power, with its Energy Market Authority naming data centres and semiconductor fabs as the demand driver, on a grid Google's own disclosure scores at 4% carbon-free. The coal extensions and the new gas are what a five-year queue does when it meets an eighteen-month build.

The tenant now carries the capacity risk

The more computable exposure is tariff duration. AEP Ohio imposed long-term capacity commitments, minimum charges on reserved-but-unused power, and collateral requirements. Tenants signed anyway, to the tune of 5,642 MW of new binding contracts by February 2026, on top of 12,219 MW contracted before the tariff. Virginia went further, putting customers above 25 MW into a separate class that pays at least 85% of contracted transmission and distribution demand regardless of usage, on a 14-year minimum term, with up to 60% of minimum charges guaranteed. Ireland's regulator, the CRU, says plainly that demand is outpacing network delivery, with data centers at 23% of metered electricity today against 5% in 2015 and a projection of 31% of national demand by 2034.

Read this as duration rather than cost. A Virginia-style contract converts flexible load into a decade-plus fixed obligation, while most neocloud and colocation revenue is contracted on materially shorter tenors. Two benign resolutions exist: tenors lengthen as AI demand proves durable, or utilities decide that enforcing minimum charges against a distressed counterparty costs them more than renegotiating. The likelier version is arithmetic anyone can run today from documents already in hand, and it is the likeliest mechanism behind a 2027-2028 failure in this asset class.

Two clocks worth carrying in diligence

The first is accounting. The GHG Protocol's July 2026 multi-statement proposal splits physical emissions from market-based emissions from corporate-action impact, and the joint GHG Protocol/ISO standard is not expected until Q4 2028. Call it two and a half years in which voluntary disclosure quality is the only differentiator. The gap that standard will eventually expose is already legible: Google discloses 100% annual renewable matching against roughly 65% carbon-free electricity measured hour by hour, Meta reports 135 tonnes market-based against 5.86 million tonnes on a location basis, and Microsoft's reported electricity emissions moved from about 260,000 to 2.7 million tonnes largely on a methodology change. Oracle claims ~92% rather than 100%, which is the most interesting number here.

The second clock is litigation, and it runs faster than climate policy. The NAACP sued xAI in April 2026 over 27 gas turbines, at least 495 MW per May 2026 injunction filings, allegedly operating without an air permit to power Colossus 2 in Southaven, Mississippi. An injunction can strand a campus mid-ramp. The investable version of speed-to-power is permitted on-site generation, so permit status belongs in environmental diligence next to carbon.

What to discount

Two rebuttals appear in every deck and both are weaker than they look. Uptime's PUE figures (1.52 average, 1.36 capacity-weighted) measure efficiency per unit of compute while absolute consumption keeps climbing. The IEA's roughly 5% of global energy-related emissions potentially avoided by 2035 is a modeled future benefit contingent on adoption, and booking it as a credit against consumption measured today is a category error someone will get paid to make.

A signed interconnection agreement is a milestone with a five-year tail, and 549 GW of them are sitting idle right now.

What to do

  1. Add three interconnection questions to every data center and power-adjacent investment committee memo starting with the next one: queue position and study stage, executed versus draft agreement, and contracted clean-generation commercial operation date against the campus energization date.

  2. Model Virginia-style terms — 85% minimum transmission and distribution demand, 14-year term, up to 60% guarantee — against each holding's actual contracted revenue tenor this quarter, and report the duration gap in years.

  3. Commission an air-permit and community-litigation scan on every portfolio site above 100 MW that runs on-site generation, using the Southaven injunction filings as the template.

The DOJ Is Now Reading Valuation Methodology

Auditors, whistleblowers and prosecutors have converged on one seam — value recognized ahead of cash — and the next limited-partner questionnaire will ask who validates your marks and what changed.

What discoverability actually changes

A markdown is an outcome. A methodology is a process with documents, dates and named owners, and processes are producible under subpoena. That is the escalation at TCP Capital: down 25% year to date, minus chief executive Phil Tseng, now facing investor litigation alongside the federal probe. The sponsor is the detail that travels, because a BlackRock-affiliated vehicle is the test case. If a marking process is discoverable at a large public business development company with an institutional parent, the argument that a private manager's process is privileged gets thin.

KPMG's finding at Guggenheim's asset-management complex is the same seam from the other side: internal-control deficiencies over how a subsidiary recognized revenue, prompted by a whistleblower, with the Financial Times framing it as trouble extending beyond Walter's insurance holdings. One auditor and one prosecutor arrived at value-recognized-ahead-of-cash in the same week. Limited partners read that coverage.

The public comp that resets defense marks

Karman Holdings (NYSE: KRMN) is the listed version of the same gap. J Capital describes a $5.3B company booking revenue growth and steady profits with no positive operating cash flow, funding serial acquisitions with debt, disclosing large deals as "not material," and carrying ongoing weak internal controls, at 70x forward earnings. J Capital is positioned in the name, so treat that as hypothesis rather than finding. The structural point survives the caveat: Karman is the ceiling comp for defense and space hardware, and a derate resets late-stage marks and exit models built on backlog rather than cash conversion.

SituationSize or multipleMechanismRead-through
TCP Capital-25% year to dateSuccessive NAV cuts, CEO exit, federal probe into valuation practicesLevel 3 process is legally discoverable; buyer discount on sponsor-reported NAV widens
Guggenheim subsidiary$367bn managerAuditor-flagged control deficiencies on revenue recognition after whistleblower complaintCounterparty and LP-capital risk where annuity-sourced commitments trace to the complex
Karman Holdings$5.3B at 70x forwardProfits without operating cash flow; debt-funded roll-up deemed immaterialCeiling comp for defense and space hardware marks

Where this collides with private AI marks

The timing is the uncomfortable part, or rather the only new part. In the same week enforcement pressure landed on mark-to-model vehicles, the private comp set printed two prices only a narrative defends. Cognition's reported raise values it at $47B, up from $26B three months earlier, on more than $900M annualized revenue, roughly 52x. Thinking Machines is reported in talks for $1B at $40B-plus on a run rate near $100M, roughly 400x. Wonderful's $550M Series C priced at $5B, up from $2B six months earlier. These are reported financing-round and in-negotiation valuations, not executable prices, and the two deserve different discount rates. One multiple has revenue underneath it. The other has a story.

The cheap parallel signal

Accounting-officer turnover is the highest-signal free indicator of a coming restatement, and it clustered: six departures across five companies worth roughly $81.1 billion in one week. Frontdoor's controller and chief accounting officer resigns September 18, after three executives sold stock in August. Amrize replaced its chief accounting officer effective September 1, days after naming a new chief financial officer. Several surfaced on LinkedIn before the filings, which is the point: the signal is free and the disclosure lag is the edge.

Managers who can answer "who validates your marks, how often, and what changed" with third-party opinions and a dated log will win allocations from managers who cannot.

What to do

  1. Commission third-party validation of your ten largest Level 3 marks and freeze a dated methodology change log before Q3 letters ship.

  2. Map every Guggenheim-affiliated exposure this week — LP commitments sourced from insurance and annuity balance sheets, NAV and subscription-line counterparties, co-invest partners — and identify a backup facility.

  3. Re-underwrite the defense and space pipeline on cash conversion this quarter: require an operating-cash-flow bridge, backlog-to-cash conversion history, and full acquisition materiality disclosure for any roll-up in diligence.

Your Portfolio's Inference Margin Was Decided at Model Selection

A 75% cache-read price cut and a near-frontier tier at one-eighth the price have both been announced, yet cost per task still rose at the capability leader.

Two kinds of numbers show up in any serving-stack write-up. The kind somebody measured under stated conditions, and the kind somebody half-remembers from a docs page they cannot find again. The vLLM claims in this passage are mostly the first kind, so they stay in. PagedAttention's near-zero cache waste stays, because that is the mechanism rather than the marketing. The 2–4x throughput over tested systems stays with the qualifier attached, since "over tested systems" is doing most of the useful work in that phrase and stripping it converts a measurement into a slogan. Automatic prefix caching stays as well, on the grounds that it is a shipped feature and not a projection. The specific context-window token count comes out. Not because it is wrong. Because nothing in the sourcing establishes it, and a precise token figure is exactly the sort of thing a reader lifts into a capacity plan and then cites back at us in six weeks when the plan does not hold. This is probably too conservative, or rather, it is conservative in a way whose cost is worth naming out loud. The number may well be accurate, somebody produces the citation, and the cut looks fussy in hindsight. Or a reader who already knows the figure reads the omission as coyness rather than discipline. Or the figure shifts, as configuration numbers tend to, and the version we would have printed ages badly while the mechanism claims do not. The view: keep the measured claims, drop the remembered one. The throughput range survives an audit. The token count does not survive one yet, and the gap between those two conditions is most of the job. The opportunity cost is real. The paragraph now says less about how much the system can hold and more about how fast it moves what it holds, which is a narrower piece of writing than the one that was drafted. It is also the piece the sourcing supports. A reader evaluating a serving stack can test a throughput multiple against their own workload in an afternoon. Nobody can test a number no one has attributed.

What to do

  1. Require cost per completed task, not cost per token, as a board reporting line at every portfolio company running agentic workloads, starting with Q3 packs.

  2. Stress-test every frontier-API-dependent position at 2x and 3x current inference pricing this quarter, and require a documented open-weight fallback and routing abstraction as a condition of the next tranche.

  3. Have portfolio counsel review the $0.10/$0.20 contributor tier for data, contribution and intellectual-property obligations before any company migrates production traffic onto it.

The $40B Youth Sports Comp Nobody Runs the Math On

One sub-segment in this sector is adding households instead of extracting from them, and the constraint on its growth is a referee shortage nobody has funded.

The demand curve, decomposed

Start with the tailwind, which is real and does not run backwards: post-2008 and pandemic-era cuts to municipal parks and recreation budgets pushed free participation into paid channels, and public fields do not come back on a five-year view. What the tailwind actually funded, though, is documented price escalation rather than more children playing. One Florida parent who paid $80 a season as a child now spends $8,000 a year on a twelve-year-old's club baseball. That same fact pattern is what makes the sector a legislative target, which is the uncomfortable part. The political risk and the return driver are the same variable.

The one category expanding the base

Girls flag football is doing what the rest of youth sports has stopped doing, which is adding people rather than repricing the ones already there. High-school participation went from under 16,000 to roughly 102,000 in four years, about a 59% compound annual rate. Reporting schools went from 2,736 to 3,990 in a single year, which is 1,254 school programs added. The 6-12 girls cohort grew 283% between 2015 and 2024, roughly 500,000 US girls aged 6-17 now play organized flag, the NFL counts more than 20 million participants across 100-plus countries, and the sport debuts at the 2028 Los Angeles Olympics.

The commercially interesting figure belongs to the Atlanta Falcons: about 50% of girls entering flag football have never played any sport before. Those are net-new households rather than wallet-share skirmishes, which means structurally lower customer acquisition cost and addressable market that actually expands. Flag is equipment-light and facility-light (no ice time, no cages, nothing resembling an $8,000 club fee), so it grows through the affordability ceiling now capping hockey, travel baseball and premium basketball.

The risk here is timing, not thesis, and it breaks a few ways. Unrivaled Sports, backed by Washington Commanders ownership, is already running a land grab and sets entry pricing before anyone else shows up. Or the Olympics pulls in enough capital that someone else's balance sheet sets it. Or the participation grows and nobody monetizes the base at all. In each version the defensible layer is relationships with the National Federation of State High School Associations, state associations and USA Football. Not facility footprint.

The bottleneck is officiating, and it is unfunded

Volume is gated by referee and umpire supply, and the shortage is explicitly attributed to the commercialized culture the sector's own economics produced. Flag football is adding over 1,200 school programs a year straight into that constraint. Recruitment, certification, assignment and scheduling software, and pay marketplaces carry non-discretionary demand from every league regardless of who owns it, earn software and services margins, need no nine-figure capital expenditure, and carry zero political exposure, because the vendor is the fix rather than the villain. Which is also why nobody is auctioning these. It is outbound sourcing work.

What to discount

Sub-segmentUnderwriting basisRegulatory exposure
Nine-figure single-purpose facilitiesARPU growth, contradicted by the one-in-five drop-off; $125M and $60M assets that do nothing elseHigh; fixed cost base, hostile politics
Exclusivity-rent mediaUp to $37 a month for streams while parents are barred from filming their own children's gamesCritical; a tying arrangement dressed as a media product
Youth-adjacent wageringOffshore Little League World Series handle reportedly exceeds WNBA and golfTerminal; Kalshi and Polymarket abstention sets the floor

There is also a pricing hole sitting in plain view: nothing between $80-a-season recreation ball and $500-$2,500-a-month premium training. The premium tier's central claim, that early specialization produces professional athletes, has no evidentiary support, and the studies point the other way on adult injury risk. A multi-sport, anti-specialization, asset-light offering at $100-200 a month is the largest unserved segment and the regulation-favored position at once.

In a sector where price is doing the growing, the only clean categories are the ones that add participants and the one that is physically out of referees.

What to do

  1. Open a girls flag football sourcing sprint this quarter across league operators, coaching networks, apparel and registration software, screening for state association and USA Football relationships as the qualifying test.

  2. Build an officiating-infrastructure thesis brief and source founders cold this quarter across recruitment, certification, assignment software and pay marketplaces.

  3. Re-underwrite hold-period and regulatory assumptions on every PE-backed youth sports position this quarter against a mandated price-transparency and non-exclusive-broadcast scenario.

The bottom line

Four unrelated markets told the same story: the number printed on the deck stopped describing the thing that is actually scarce, and in each case an outside party — a national laboratory, a prosecutor, an open-source default, a legislator — did the reclassifying rather than the market. That breaks the habit of underwriting a headline metric because it is the one the seller chose to report. The defensible position is owning the denominator: the energized asset, the cash conversion, the completed task, the net-new participant. Restate your ten largest positions this week in the units their real constraint is measured in, and see which theses survive the translation.