Venture Math · Fall 2026

Warrants & Strategic Investors

Session 06 · Oct 7 · Andy Ku (AI Fund, ex-IQT)
Data: SEC filings, company announcements, Strebulaev & Wang · as of Oct 6, 2026
The strategic’s side — what corporates put in, and what they take backCLAY = WHAT THE STRATEGIC TAKES

In AI, the investor is also the supplier and the customer

Sized by market cap (Oct 6) or last round · “up to” amounts as announced · filings and releases; layout after Bloomberg

Chips or hardwareInvestmentCloud and servicesWarrants and guarantees

Web of AI deals among OpenAI, Anthropic, Nvidia, Microsoft, Amazon, Alphabet, AMD, Meta, Oracle, CoreWeave and SpaceX/xAI

Amazon’s warrants: small caps give up ~20%

Warrant as % of shares outstanding; vests as Amazon’s spend hits the cap at right. Nothing between 4% and 15%.

Plug Power '17 28.9% $600M Clean Energy '21 26.7% $500M ATSG '18 25.1% vested StarTek '18 24.8% $600M Ranpak '25 22.4% $400M Velodyne '22 20.2% $200M Hawaiian '22 18.4% $1.8B Applied Opto '25 16.0% $4B SpartanNash '20 15.2% n/d TD Synnex '26 4.0% n/d Generac '26 2.9% $8B STMicro '26 2.8% milestones Plug Power '22 2.8% $2.1B Qualcomm '26 2.4% $60B Astera Labs '26 1.9% $6.5B Jabil '24 1.0% n/d Marvell '24 0.5% to 2030

8-K / 10-K warrant terms · 17 Amazon grants, 2017–2026

<4%
of CVC-backed startups are later bought by a corporate backer. When a parent does buy, its own shareholders lose value. PitchBook Q1 2025 · Benson & Ziedonis

The check is the small part of an AI megadeal

Equity in vs. spend contracted back to the investor. Median 6×.

Microsoft → OpenAI $13B in $250B back 19.2× Nvidia → xAI SPV $2B in $20B back ~10× Microsoft → Anthropic $5B in $30B back 6.0× Amazon → Anthropic $25B in $100B back 4.0× Amazon → OpenAI $50B in $100B back 2.0×

Microsoft & Amazon 10-Qs, company releases · Nvidia–xAI reported only

What strategics ask for that VCs don’t

% of 74 S&P 500 CVC units that often or always require it.

Pro-rata rights 92% same as a VC Observer seat (required or preferred) 77% 31% require it Follows rather than leads 57% lead VC sets price Extra rights by side letter 55% Right of first notice on a sale 50% Commercial deal as a condition 27% Right of first refusal on a sale 12% 0% in California Full board seat required 4% 38% never take one

Strebulaev & Wang, table 17

1 · CoverageLender: warrant $ as % of the loan. Customer: shares as % of the company.
2 · StrikeLender: last round. Amazon: 30-day VWAP. Government, AI compute: often $0.01.
3 · VestingCustomer: as it spends with you. Lender: at signing and on draw.
4 · ExerciseCash brings money in. Cashless brings none and issues fewer shares.
The other strategics — chip suppliers, the CIA’s venture arm, studiosCLAY = STRATEGIC

In AI compute, the supplier pays the customer in stock

Grantor → holderWarrantWhat earns it
AMD → OpenAI · Oct ’25160M sh at $0.01 (~9.9%)First tranche at 1 GW of GPUs shipped, all at 6 GW, plus AMD price targets up to $600
AMD → Meta · Feb ’26160M sh at $0.01 (~9.8%)Same 1 GW / 6 GW structure and price hurdles
Marvell → Google · Aug ’2659.0M sh at $206.58 (6.7%)240 tranches, one per $500M of custom-chip revenue: $120B to vest fully
Cerebras → OpenAI · Dec ’2533.4M sh at $0.00001Funding a ~$1B working-capital loan and taking 750 MW (up to 2 GW)
TeraWulf → Google · Aug ’2573.6M sh at $0.01 (18.8%)Google backstopping Fluidstack’s data-center leases. No purchases at all
Cipher → Google · Sep ’2524.2M sh at $0.01 (6.1%)Lease backstop. Cipher tops up if the shares are worth under $430–435M

8-Ks and 10-Qs · AMD: no tranche vested as of Jun 27, 2026

IQT runs on ~$110M a year of government money

Federal contributions to In-Q-Tel ($M, fiscal years ending March). Below: IQT’s own claim of VC dollars that followed each IQT dollar, up ~3× in four years.

$0M $50M $100M 128 FY17 104 FY19 116 FY21 $22 99 FY22 $28 118 FY23 $32 103 FY24 $40 114 FY25 $62 VC $ per IQT $:

IQT Form 990s via ProPublica · the multiplier is IQT’s statistic, not a causal estimate

The work program shows up in the tax return

IQT’s biggest contractors are its own startups, paid to adapt their product for government users. ~200 contractors a year get >$100K.

StartupTax yearPaid by IQT
MemSQL2016$3.95M
OmniSci2020$4.25M
Chainalysis2020–21$3.32M
Orbital Insight2016$2.42M
Matterport2021$1.53M
Workera2024$1.34M
GitLab2018$1.23M

990 Part VII-B · “technology development”

What IQT’s portfolio became

CompanyIQT entryOutcome
Keyhole2003, ~$527K roundGoogle Earth (2004)
Palantir~2005, ~$2M (rep.)Direct listing 2020; ~$455B
FireEye2009IPO 2013
Recorded Future2009, <$10M roundMastercard, $2.65B (2024)
Kensho~2016S&P Global, ~$550M (2018)
Apigeepre-2015Google, $625M (2016)
Elementalpre-2015AWS, ~$296M (2015)
GitLabby 2018IPO 2021

IQT releases, 990s, SEC filings, press · no study shows IQT money causes more follow-on

Studio founders own about half as much

Modeled cap table after seed, by path. Same seed round (19.5% sold) and 12% pool in each.

Founder + SAFE + seed fund 61% 20% 12% YC + seed fund 62% 20% 12% Studio (GSSN average split) 36% 33% 20% 12% Majority studio (eFounders 2015) 34% 34% 20% 12% Carta median founding team after seed: 55.8%

Venture Math model on Carta medians · GSSN 2020 split (studio 34 / founder 50 / pool 16)

Notable studio exits, and what the studio kept

CompanyStudioOutcomeStudio stake
ModernaFlagshipIPO 201817.9% after IPO
Generate:BiomedicinesFlagshipIPO Feb 2026, $400M48.8% after IPO
Hims & HersAtomicSPAC Jan 2021, ~$1.6B5.9%
Dollar Shave ClubScience Inc.Unilever, ~$1B (2016)n/d
OvertureIdealabYahoo, $1.63B (2003)n/d
WorkeraAI FundPearson deal, Sep 2026n/d

And the misses: eToys (Idealab; IPO 1999, bankrupt 2001) · Idealab’s own IPO (shelved 2000) · Rubius (Flagship; IPO 2018, liquidated 2023) · Woebot’s consumer app (AI Fund investee; shut 2025)

SEC prospectuses and proxies (Moderna, Generate, Hims) · Fortune, Idealab, Pearson, BioSpace