The money goes in circles.
So do the chips.

A dozen companies are building the AI economy by investing in each other and then buying each other's products. Nvidia funds OpenAI, which buys Nvidia chips. AMD hands OpenAI a slice of itself in exchange for orders. Microsoft, Amazon and Google pour billions into model labs that pour it back into their clouds. This is a map of that loop: the capital, the hardware, the energy it burns, and the question everyone is asking: flywheel, or hall of mirrors?

The Tape
The Money Loop 01 / 06

Tap a company. Watch where its money goes, and what comes back.

Blue is investment: equity, warrants, capital flowing in. Teal is spending: chips, cloud and compute flowing back. When the two close a circle, you are looking at a loop. Dashed lines are letters of intent; dotted claret lines have stalled.

Deal status
Invests in Spends back Backstop Supplies
Show
↔ drag / scroll to explore
Follow the money.
Tap any company in the map to see who funds it, what it spends, how much, and whether each deal is confirmed, merely reported, a non-binding letter of intent, or stalled.
Follow the Dollar 02 / 06

Three loops that explain the whole machine.

Strip away the noise and the same shape keeps repeating: capital goes out, product orders come back, and both sides book bigger numbers.

01

The headline loop

Nvidia ⇄ OpenAI

Nvidia signalled it could put up to $100B into OpenAI as data centers come online. OpenAI commits to 10+ GW of Nvidia systems. The chipmaker funds its own biggest customer.

Letter of intent, not yet signed
02

Equity for orders

AMD ⇄ OpenAI

AMD handed OpenAI a warrant for up to ~10% of itself at a cent a share, vesting only as OpenAI deploys 6 GW of AMD chips and AMD's stock climbs. It later struck the same deal with Meta.

Confirmed: warrant tied to milestones
03

Cloud round-trip

Hyperscalers ⇄ Labs

Amazon and Google invested billions in Anthropic; Anthropic commits $100B+ back to AWS and $30B to Azure. The twist: Alphabet booked a $36.9B paper gain just marking up its own stake.

Confirmed, plus large paper gains
Bubble or Boom 03 / 06

Two camps reading the same deals.

The circular financing looks like dangerous self-dealing to one side and the first innings of the largest infrastructure build in history to the other. Here is the honest version of each case.

The Bears ↓

"It's a hall of mirrors"
Bank of England
Financial Policy Committee

Warned the risk of a sharp correction has risen, called AI valuations particularly stretched, and noted the top five S&P 500 names are a bigger share of the index than at any point in 50 years.

Michael Burry
Scion Asset Management

Disclosed puts against Nvidia and Palantir and accused hyperscalers of understating depreciation, stretching chip lifespans on paper to flatter earnings despite an 18–24 month product cycle.

Academic economists
CEPR / VoxEU

Argue circular funding means the same capital appears several times across different balance sheets, creating hidden fragility, and that the real danger arrives if bank credit enters the loop.

IMF leadership
Georgieva & Gourinchas

Told markets to buckle up and drew explicit dot-com parallels, flagging that financing is quietly shifting toward debt and private credit.

Harris Kupperman
Praetorian Capital

Runs the depreciation math: one year of AI capex needs roughly $480B of revenue to earn a fair return, against $15–20B realized today. A ten-fold gap that closes only if growth stays vertical.

MIT Project NANDA
State of AI in Business 2025

Found that about 95% of enterprise AI pilots show no measurable return yet, with only ~5% extracting real value. If that holds, the revenue to justify the buildout never shows up.

The Bulls ↑

"A 1996 moment, not 1999"
Jensen Huang
Nvidia, Chief Executive

Rejects the bubble framing (we see something very different), calling it the largest infrastructure build-out in human history, with the world only a few hundred billion into the trillions required.

Dan Ives
Wedbush Securities

Calls it a 1996 moment, not 1999, and estimates every $1 of AI hardware spend generates $8–$10 of downstream ecosystem value.

Goldman Sachs
Research

Projects roughly $1.15T of hyperscaler capex across 2025–27 and argues cash flow plus investment-grade credit could fund up to ~$1.4T, with markets rewarding a clear capex-to-revenue link.

The IMF's other half
Chief economist

Concedes a bust is possible but notes the boom is largely equity-financed, so unlike the dot-com era a correction does not necessarily transmit to the broader financial system.

Satya Nadella
Microsoft, Chief Executive

"Jevons paradox strikes again." As inference costs fall roughly ten-fold a year, and DeepSeek's R1 undercut comparable models by ~96%, cheaper intelligence expands demand far faster than it deflates revenue.

Enterprise adoption
Menlo Ventures / Anthropic

Paid, fast and real: Anthropic's run-rate went from ~$1B to ~$47B in about 18 months, and it now holds ~40% of the enterprise LLM market against OpenAI's 27%. Demand, not vapor.

What would settle it: seven signals to watch
Capex vs realized revenue

Bear: the gap keeps widening. Bull: revenue compounds faster than capex.

Model-lab gross margin

Bear: stuck at 30–50%. Bull: marches toward 70%+ as inference deflates.

Financing terms

Bear: down-rounds, vendor financing, junk-rated debt. Bull: up-rounds, investment-grade issuance.

Datacenter utilization

Bear: stranded or idle capacity. Bull: demand outruns supply and rate limits bite.

GPU depreciation

Bear: a hyperscaler shortens useful life or writes down. Bull: useful lives hold and a value cascade appears.

Power delivery

Bear: megawatts gate revenue while capex is sunk. Bull: turbines, grid and nuclear keep pace.

Enterprise AI ROI

Bear: the 95%-no-return persists. Bull: documented P&L gains broaden beyond tech.

The Other Circular Economy 04 / 06

The loop that's literal: metal, water, and watts.

Circular economy also means the real thing: reusing the hardware, recycling the water, and powering it all. The hardest constraint is no longer chips, it is power, and this loop runs behind the buildout it has to keep up with.

Hardware reuse
90.9%
Microsoft server reuse + recycle rate, FY24

Hyperscalers run tight hardware loops. Oracle reports 99.6% reuse, Google has resold 44M+ components since 2015, and refurbishing can save up to 85% of manufacturing energy.

But globally only 22.3% of e-waste is properly recycled, and just ~28% of operators even track decommissioned gear.
Power & grid
$163B
GE Vernova power-equipment backlog

The binding constraint moved from chips to electricity. Gas-turbine slots are booked into 2029–30, transformers are scarce, and interconnection queues run years. Every major lab has also gone nuclear: Microsoft is restarting Three Mile Island and Meta has lined up ~5–6 GW of reactors.

The catch: most turbines, transformers and reactors will not deliver until 2029–2030+, well after the demand arrives. Backlogs prove the demand is real (ASML €38.8B, Broadcom $73B), but the megawatts arrive late.
Water
of new data centers sit in water-stressed areas

A big data center can draw up to 5M gallons a day. Texas centers were on track for ~49B gallons in 2025, potentially ~399B by 2030. New zero-water cooling designs aim to cut that sharply.

Critics call the green messaging overstated: U.S. residential power prices rose 7.1% in 2025, double the rate of inflation, as the grid absorbs the load.
The Connected Companies 05 / 06

Everyone in the loop, and where they trade.

Prices are snapshots from late June 2026 and move constantly, so treat them as a sense of scale, not a quote. Private valuations reflect the latest funding rounds.

CompanyTickerCategory~Price~Mkt capRole in the loop

Private · OpenAI ~$852B (IPO filed) · Anthropic ~$965B · xAI ~$230B · Stargate >$400B committed · SoftBank a leveraged proxy for OpenAI

Your Portfolio Is in the Loop 06 / 06

You probably own this bet without choosing it.

The same dozen names sit at the top of the index funds and pensions almost everyone holds, which is exactly what the central banks worry about.

Seven companies now make up a third of the entire S&P 500. The other 493 share what's left.

~38%
of the SMH chip ETF is just Nvidia + TSMC + Broadcom
~22%
of the Magnificent 7's weight is Nvidia alone
43 days
for the DRAM memory ETF to reach $10B, a record

A standard S&P 500 or Nasdaq-100 (QQQ) fund now embeds roughly a one-third AI-concentration bet. Sector ETFs like SMH and SOXX do not diversify that bet. They multiply it. Equal-weight funds such as RSP are one of the few ways to dilute it.

And you may already own the private labs

OpenAI, Anthropic and SpaceX are private, but their largest backers are public mega-caps almost everyone holds. Each stake is a low single-digit to mid-single-digit slice of the parent, so the optionality rides nearly free on top of a real, cash-generative core.

Microsoft → OpenAI
~$135B
~27% stake, about 4.5% of Microsoft's market cap

Plus a $250B Azure backlog. The revenue share is now capped at $38B and ends at AGI, and the model license is no longer exclusive.

Alphabet → Anthropic + SpaceX
~$135B + ~$87B
~14% of Anthropic (15% cap, nonvoting) and ~5% of SpaceX

Together roughly 4.5–5% of Alphabet's cap, on top of Search, YouTube and Cloud. The cleanest single proxy for both.

Amazon → Anthropic
~$74–180B
convertible notes + nonvoting preferred, re-marked each round

Higher torque, more complex structure. Amazon booked $9.5B and then $16.8B in gains as Anthropic's mark climbed.

A reported "%" stake is not clean economic value: capped, nonvoting and conversion-dependent structures change what you would actually realize. And the premium retail wrappers are best avoided, VCX traded near a ~390% premium to NAV and DXYZ ran to ~2,000% before falling 80%+. Listed funds like Scottish Mortgage, ARK Venture and XOVR are cleaner routes.
Pass it on

Help the map travel, and help it get better.

This is a living picture built from public reporting. Send it to someone who should see it, and tell me what I have missed.

Spread the map

One link, the whole picture. Post it, message it, or drop it in the group chat.

Post on X inLinkedIn WhatsApp Reddit Email

Improve the map

Spotted a deal that is wrong, missing, or out of date? Have a better source? That is exactly what keeps this accurate.

Corrections reach a person, not a database. Add a link to your source if you can, and say which deal it concerns.

Methodology & sources

How this was built, and where every number comes from.

This is desk research, not original reporting: public company announcements, SEC filings, central-bank notes and major-outlet coverage, assembled into one picture. Every deal below links to its source.

What the status labels mean

Confirmed is a signed, disclosed deal. Reported is credible reporting, not formally confirmed. Letter of intent is a non-binding commitment. Stalled means announced, then reportedly paused. The label reflects a reading of the latest reporting, not the linked source alone.

What is dated, not individually cited

Share prices and market caps are snapshots as of June 2026 and move constantly, so they are dated rather than separately sourced. Private valuations come from the latest funding rounds. Forward-looking figures are flagged as estimates. The optional supply-chain layer ("All companies") shows well-established structural relationships, such as who fabricates whose chips, as context rather than individually sourced deals.

Every deal, with its source

FlowDetailStatusVerifiedSource

Read this before you share it Methodology

  • Reported is not confirmed. Many headline numbers (Nvidia's $100B, Oracle's ~$300B, the $250B Azure commitment, Broadcom's custom-chip phase) are letters of intent or multi-year commitments, not cash already spent. Several were explicitly unsigned or stalled in early-to-mid 2026.
  • Prices are snapshots. Share prices and market caps were observed in late June 2026 and change every second. Sources sometimes disagree on a single company's market cap. Do not trade off these figures.
  • Private valuations are round marks. The OpenAI, Anthropic and SpaceX/xAI numbers come from their latest funding rounds and may include conditional or compute-credit components, not liquid market value.
  • Run-rate is not revenue. Headline annualized run-rate figures (OpenAI's ~$20B+, Anthropic's ~$47B) take a single strong month and multiply by twelve. Full-year realized revenue is far smaller, and the gap worth tracking is 2026 capex against realized revenue, not against run-rate.
  • Revenue is counted differently. Anthropic books cloud-reseller revenue on a gross basis; rivals argue the comparable net figure is lower. Treat cross-company revenue and margin comparisons as contested, not settled.
  • Forward-looking items are predictions. Capex forecasts, reactor delivery dates, IPO valuations and analyst targets are estimates, presented as such.
  • This is a map, not advice. It is built from public reporting to help you see the shape of the system and judge the debate yourself. It is not investment advice.