ClarityBeta

Investment & Market Intelligence

The Investor

7 min read · 1,307 words · 3 moves to watch
The Signal

CalPERS cut buyouts from 91% to 58% of new commitments in three vintage years.

The rotation took that PE program from 30th to 1st among the 30 largest US pensions, which is the kind of result peers copy, and MassPRIM and North Carolina are copying it. The same dataset carries the awkward part: only 17.0% of 2,143 funds from 2000–2018 vintages returned 2x DPI. Which means the peer benchmark you get measured against is rotating toward a product five in six managers never delivered.

In Play

  1. Discount Rate Rises Into Next Week's Fed Meeting

    The 10-year Treasury closed at 4.944%, up 11 basis points, per Morning Brew, and fed-funds futures now price roughly a 70% chance of a quarter-point increase at the September 15-16 meeting. Wholesale inflation printed in line with expectations and investors sold anyway, which says the market is pricing fiscal and energy risk rather than a monthly data point. Your live deal models and Q3 marks are struck against a lower risk-free rate than the one being quoted, and terminal-value-heavy models take the largest hit.

    Ask Clarity
    Try

  2. Financings That Set No Price At All

    Motive withdrew its IPO registration on Thursday and disclosed more than $1.3 billion from General Catalyst's Customer Value Fund at an undisclosed price, per Paul Smalera's private-market reporting. Stoke Space's $1 billion Series E first close and Cylake's $245 million convertible note also landed with no equity reference. Each one is a stale mark in formation, and the deferred price discovery that produced this cycle's steepest software resets is being manufactured again in these deals.

    Ask Clarity
    Try

  3. Power Supply Now Gates AI Capacity

    Microsoft CFO Amy Hood said "very little can get built and come online in the next 12 months" while committing to grow from 12GW of capacity to over 38GW by 2032, per The Information's reporting. Bloomberg Technology notes that 38GW exceeds New York State's entire projected 33GW peak summer demand. Google paired a $15 billion Finnish buildout through 2028 with a 22-year Fortum power contract. For you, AI input costs stay rationed and flat-to-rising through H2 2027, which breaks any app-layer margin model built on falling inference prices.

    Ask Clarity
    Try

  4. CalPERS Rotated 34 Points Into Growth and Venture

    CalPERS cut buyouts from 91% to 58% of new commitments and lifted growth equity plus venture from 9% to 43% between FY2020-21 and FY2023-24, moving its private equity program from 30th to 1st among the 30 largest US pension programs in three years, per a16z. MassPRIM and the North Carolina Treasurer are reported to be following. That is a documented allocator rotation you can raise into. The same data set undercuts the easy conclusion: of 2,143 global funds from 2000-2018 vintages, only 17.0% returned 2x DPI.

    Ask Clarity
    Try

  5. Physical-AI Series A Rounds Now Clear at $100M

    Fortune's Term Sheet counted roughly $749.9 million across ten venture rounds, and two physical-infrastructure deals took about 53% of it: Celero's $275 million Series C for moving data between AI data centers and TAR's $120 million Series A for off-grid data-center power. Maven Robotics raised a $100 million Series A for industrial robots. Analog Devices agreed to pay $1.35 billion cash plus a $200 million earnout for Alif Semiconductor. If your ownership model in this category assumes $15-25M A rounds, it no longer works.

    Ask Clarity
    Try

Deep Dives

Three of the Biggest Financings Set No Price At All

A withdrawal ends a registration rather than pausing one, and the private capital replacing it leaves no reference price behind for the holders who will need one.

What a withdrawal actually does

Motive filed for an NYSE listing on December 23, 2025, delayed in January, and on Thursday filed a Form RW. The mechanical difference is the whole story: a delay leaves the registration statement on file, while a withdrawal ends the process and requires an entirely new filing to restart. There is no revised timetable, no disclosed valuation, and no priced reference. In its place came more than $1.3 billion from General Catalyst's Customer Value Fund at an undisclosed price, with Pranav Singhvi joining the board and CEO Shoaib Makani framing it as staying private while continuing to invest.

This is not a rescue. Motive carries roughly $600 million of ARR growing 30% year over year, with revenue from six-figure customers up nearly 60%, per Paul Smalera's reporting. A company with that profile chose structured private capital over a public price, which tells you what the sponsor thinks a public price would have been.

Zero reference prices, three deals

Motive was not alone. Stoke Space announced a first close on a $1 billion Series E with valuation undisclosed, taking total funding to $2.3 billion. Cylake raised a $245 million convertible note six months after a $45 million Greylock seed, while still pre-beta. That is nine figures of capital, three times over, with no equity reference attached.

The mechanism that produced this cycle's ugliest software resets was never business deterioration. It was deferred price discovery: companies grew, never took another priced round, and carried a peak mark as the official reference until a transaction forcibly replaced it. Every one of these unpriced financings recreates that condition for a new cohort of holders.

Why the timing is different from six months ago

The 10-year Treasury closed at 4.944%, up 11 basis points, per Morning Brew, and wholesale inflation printed in line while investors sold anyway. Fed-funds futures price roughly a 70% chance of a quarter-point increase at the September 15-16 meeting. Not a cut. Late-stage private marks respond to discount rates with a lag, which means the repricing pressure now arriving has not yet passed through the book.

The divergence between the two markets is measurable in the same period. Boring Co. raised at a $23 billion valuation while the Nasdaq closed lower for a fourth consecutive session. And where a public price does exist, it is unsentimental: SpaceX closed at $147.55, down 3.86%, only about 9% above its $135 IPO price, with 328.4 million shares unlocking September 24, further tranches on October 9 and 24, and up to 1.3 billion tied to Q3 results.

Where the sources disagree

a16z reads the two-speed market as asset-class strength: trophies trade at premiums to last round, and actively-raising unicorns compressed median time between rounds to 1.0 year in Q1 2026 from 1.5 years in 2024 (Mach Industries went $1.8B to $3.7B in three months; Ramp is in talks at roughly $60B against a $44B Series F). The private-markets reporting reads the same distribution as deferral, noting that 51.2% of companies that once cleared unicorn status have not raised in over two years. Both descriptions are accurate. Only one of them describes the half of the market where most fund NAV actually sits.

A withdrawn registration is not a slower IPO. It removes the only mechanism that would have printed a price.

The practical consequence is that headline equity value and common recovery have decoupled, and structured growth capital sits between them. Waterfall first, mark second.

What to do

  1. Re-run every top-ten position and live deal model at a 5.0% risk-free rate this week and publish the NAV sensitivity in the Q3 LP letter before quarter-end.

  2. Commission a liquidation-preference waterfall on every position whose last reference price came from an undisclosed or structured financing, starting with Customer Value Fund-style growth deals, by quarter-end.

  3. Require a written price-discovery milestone in every new structured or convertible term sheet you sign from this quarter forward.

Microsoft Says Almost Nothing New Comes Online For Twelve Months

The demand-side rationing that portfolio companies have been absorbing since April turns out to be a supply statement, and it fixes the input cost curve through H2 2027.

The rationing came first

The behavioral evidence has been visible for months and was misread as pricing power. OpenAI paused new $200-per-month Astra subscriptions citing unprecedented demand and system strain. Anthropic barred $20-to-$200 Claude tiers from powering third-party agents back in April. AWS raised AI workload rental prices 20% in June. Frontier vendors are turning away revenue, which no growth-stage company does voluntarily.

Amy Hood supplied the reason, per The Information's reporting: "Very little can get built and come online in the next 12 months." Microsoft cannot increase planned capacity for at least a year and is instead squeezing efficiency from existing servers, while committing to grow from 12GW to over 38GW by 2032. It has been turning away customers who want AI-chip servers and at times chose between growing Azure and reserving capacity for Copilot. It has largely run out of grid-supplied power, and is renting from CoreWeave, Nscale, Nebius and rival AWS. Bloomberg Technology's framing sizes the ambition: 38GW exceeds New York State's projected 33GW peak summer demand.

Twenty-two-year offtakes funded by six-month leases

Google's largest European commitment, $15 billion through 2028, was announced alongside a 22-year Fortum power agreement and stated intent to explore new reactors at Loviisa. That is the duration a serious compute footprint now requires. On the other side of the same market, SpaceX signed a compute-landlord deal at roughly $1.11 billion per month starting December 1, on top of Google at $920 million per month, and its CFO says "almost all" such contracts run about six months.

That mismatch is the diligence question for anything neocloud-adjacent in your pipeline. Long-dated physical obligations are being serviced by rolling short-term revenue against a handful of counterparties, and the renewal cliff is not priced in most models.

Where the sources genuinely diverge

Oracle is the test case, and the reporting does not agree on the verdict.

DimensionDisclosed figureWhat it supports
Cloud infrastructure revenue+121% to $7.4BDemand is real and converting
Remaining performance obligations$664B, up $209B y/yBacklog underwrites the neocloud debt stack
Quarterly capex$28.5B, FY plan held at $90-95BMulti-quarter visibility, irreversible cost base
Interest expense+55% to $1.4B, excluded from operating incomeThe financing cost is not in the headline margin
Market reaction~+7% after hours; -5.4% to $152.94 in Bloomberg Technology's account of the sessionReports diverge on whether capex-bought revenue earns a premium

Read together, both readings hold: demand is not in question, and the equity is now being underwritten on return on invested capital rather than growth. MediaTek's 44% monthly sales surge on AI chips for customers including Google says the same thing from the silicon side — if a portfolio company's gross margin is a function of GPU scarcity, that margin has a shorter half-life than the model assumes.

A 22-year power contract serviced by six-month compute leases is a duration mismatch, not a moat.

The actionable read for the app layer is narrower and harder: underwrite inference input costs as flat-to-rising through H2 2027, not declining. Any company whose gross-margin ramp depends on cheaper tokens next year is modeling a supply curve that Microsoft has publicly said does not exist.

What to do

  1. Add a mandatory power section to every AI-infrastructure investment memo before your next committee: interconnect queue position, firm-power cost per MWh, PPA duration, and behind-the-meter or nuclear optionality.

  2. Commission gross-margin sensitivity work this quarter on every portfolio company with more than 30% frontier-API dependency, modeling flat-to-plus-20% inference pricing through H2 2027.

  3. Re-underwrite compute-rental, neocloud and data-center exposure for contract duration and counterparty concentration before the next valuation cycle, pricing the six-month renewal cliff explicitly.

The Buyout Exit And The Seat-Pricing Exit Are The Same Trade

Allocators are stepping away from levered software just as the vendors themselves concede that the seat is no longer the unit customers will pay for.

Start with the credit side, not the allocation. Buyout returns — average and top quartile — sit near 15-year lows, per a16z's own scorecard. The franchise premise behind a decade of software LBOs, Robert Smith's argument that software contracts are better than first-lien debt, is being undermined by AI making recurring revenue less durable than underwritten — with the $50,000 figure removed.

What to do

  1. Screen the book by quarter-end for indirect levered-software exposure: direct credit, sponsor-backed equity, and portfolio-company ARR concentrated in PE-owned software customers walking into refinancings.

  2. Require every application-software position whose moat memo cites migration friction to present a non-lock-in defensibility case — data custody, workflow depth, distribution — at its next board meeting.

  3. Rebuild the LP narrative around the documented pension rotation and name the allocators still sitting at 5-10% venture targets before this becomes standard GP boilerplate.

The bottom line

The pattern across these items is that every input to a private valuation moved in the same period while the valuations themselves did not: the discount rate, the cost and availability of power, and the durability of subscription revenue. That retires growth as evidence of durability — a position is now only as sound as the input it is most exposed to and the transaction that last proved its price. Grade every mark in the book this week by which of those three inputs moved against it, and treat any reference price resting on an unpriced financing as unpriced until a real transaction says otherwise.