⛵Dhow Dispatch │ The Circular AI Economy

Download Dhow - the community-powered private markets platform helping Muslim investors discover opportunities, learn, connect, and back Muslim founders building the future.

Our $10,000 allocation spots are fully filled, and Dhow Horizon Fund I has now made its first investment in Lore, an AI-powered entertainment intelligence platform. The minimum commitment for new investors is now $25,000, and access will become more limited as we continue deploying the fund. Reserve your allocation here.

A strange debate broke out this week around Harvey.

Harvey is one of the breakout companies of the AI application boom. The legal AI startup raised another $550M on September 9 at a $15.5B valuation and says more than 80% of the Am Law 100 now use its software. The company has also crossed $400M in annual recurring revenue.

That puts Harvey at roughly:

$15.5B valuation ÷ $400M ARR = ~39x ARR

Then another number emerged.

Earlier this year, Harvey rolled out more agentic products and usage exploded. Bloomberg reported that its gross margin fell from roughly 50% at the beginning of the year to -50% by June as model costs surged. Harvey co-founder Gabe Pereyra later said the company brought margins back above zero within a quarter through routing, post-training, infrastructure improvements, and better cost controls, even while usage continued growing rapidly.

A -50% gross margin has a simple implication.

For every $1 of revenue, Harvey was temporarily incurring roughly $1.50 of cost of revenue.

Now look one layer below Harvey.

Anthropic raised $65B at a $965B post-money valuation in May, when the company reported a $47B revenue run rate. That works out to about 20.5x run-rate revenue.

So take a stylized $1M Harvey customer contract:

$1M of Harvey ARR × ~39x = ~$39M of Harvey valuation

At the June -50% gross margin snapshot, $1M of revenue implied roughly $1.5M of cost of revenue. If, purely for illustration, that entire amount became Anthropic revenue:

$1.5M × ~20.5x = ~$31M of additional valuation

Add them together:

This is an illustration, not literal accounting. Harvey's cost of revenue includes more than Anthropic, its margins have already recovered, and its model mix is changing as it adopts open-weight models. The arithmetic exposes something more useful: the same underlying pool of enterprise demand can support highly valued revenue streams at multiple companies simultaneously.

And Anthropic still has to buy compute.

Harvey becomes more interesting when you put it beside the rest of the application layer.

Its closest competitor, Legora, says it passed $200M ARR last week. It took the company 18 months to move from $1M to $100M, then less than six months to add the next $100M. Legora says 130,000 lawyers now use the platform monthly across 2,100 firms and legal teams.

At the same time, Legora is reportedly discussing a raise of at least $300M at roughly an $8.5B pre-money valuation, only months after closing its Series D and extension at about $5.6B. The proposed terms would place it at approximately 42.5x its newly reported ARR. The financing remains under discussion.

Nvidia's venture arm is already on Legora's cap table.

The same aggressive capitalization appears elsewhere:

*Glean's $7.2B valuation dates to June 2025 while the $300M revenue milestone came in May 2026, so this is a reference ratio rather than a round-date multiple. TechCrunch also noted that some of Glean's consumption revenue makes its $300M figure less comparable to traditional subscription ARR.

Perplexity belongs in the same conversation. Its annualized revenue reportedly climbed from less than $250M at the start of 2026 to more than $750M, while Nvidia has reportedly discussed investing at a valuation above $30B. Perplexity also signed a major Azure commitment earlier this year.

These companies are generating substantial revenue. Enterprise demand is visible.

The financial question begins when we follow that revenue upstream.

Imagine a company spends $1 million on an AI application.

The application provider books revenue.

It then pays a model provider to process prompts, reason over documents, generate code, or operate an agent.

The model provider pays for enormous quantities of compute.

That compute runs through cloud providers and specialized AI clouds, which buy chips, networking equipment, power, and data-center capacity.

The chain looks roughly like this:

↓

Harvey, Legora, Cognition, Sierra, Lovable, Glean
↓

Anthropic, OpenAI and others
↓

AWS, Azure, Google Cloud, CoreWeave and other providers
↓

Nvidia, Broadcom, data centers, power, networking

Suppose the original business spends $1.00.

The application company could record that $1.00 as revenue. It might send $0.40 upstream to model providers. Those providers could spend part of that $0.40 on cloud capacity. The cloud operator uses some of its revenue to buy accelerators and build data centers.

Across the chain, the original dollar can contribute to considerably more than $1 of gross reported revenue across different companies.

That is standard value-chain accounting.

A restaurant's $100 dinner creates revenue for the restaurant, its food distributor, the farmer, the logistics company, and the landlord. Nobody adds those figures together and calls the total independent consumer demand.

AI investors need to apply the same discipline.

That distinction matters more as the companies sitting at each layer receive very large valuations based on their respective revenue streams.

The AI economy becomes more unusual when money moves in the opposite direction.

Infrastructure suppliers are investing in customers.

Cloud providers own pieces of model companies that commit to spending enormous amounts with those same cloud providers.

Chipmakers are providing capital and guarantees to businesses whose growth creates demand for more chips.

Microsoft has made $13B in total funding commitments to OpenAI and holds roughly 27% of OpenAI's PBC on an as-converted diluted basis.

OpenAI, meanwhile, has contracted to purchase an additional $250B of Azure services.

The loop is straightforward:

Microsoft capital → OpenAI → Azure purchases → Microsoft cloud revenue

The transactions are contractual and economically substantive. They also connect Microsoft's investment return to the growth of one of Azure's enormous customers.

Amazon's relationship with Anthropic is even more intertwined.

Anthropic has committed to spend more than $100B on AWS technologies over ten years, securing as much as 5GW of compute capacity. Amazon invested another $5B when the expansion was announced and said up to another $20B could follow.

Amazon's June 2026 10-Q adds another layer. The company disclosed a financing facility for Anthropic that can reach $20B, with availability tied to Amazon meeting compute-capacity delivery milestones. Amazon also recorded $50.5B of upward valuation adjustments in Q2 alone on its Anthropic preferred stock after new financing rounds reset the observable value of the stake.

Read those relationships together:

Amazon invests in Anthropic
↓
Anthropic commits >$100B to AWS
↓
AWS builds capacity and recognizes revenue over time
↓
Anthropic grows and raises at higher valuations
↓
Amazon marks up its Anthropic investment

This is a powerful flywheel when demand continues compounding.

Anthropic reportedly committed to spend $200B with Google Cloud over five years, while Alphabet has separately committed to invest up to $40B in Anthropic.

Reuters reported that Anthropic alone could represent more than 40% of Google's recently disclosed cloud backlog under the terms reported by The Information.

There is another Anthropic loop.

Microsoft agreed to invest up to $5B in Anthropic. Nvidia agreed to invest up to $10B.

Anthropic simultaneously committed to purchase $30B of Azure compute capacity and contract additional capacity using Nvidia architecture.

The investors are also important suppliers.

Nvidia and OpenAI announced an agreement to deploy at least 10GW of Nvidia systems, representing millions of GPUs.

Nvidia said it intends to invest up to $100B in OpenAI progressively as each gigawatt is deployed.

The capital commitment is therefore directly connected to infrastructure deployment.

OpenAI signed an initial $11.9B compute contract with CoreWeave.

As part of the transaction, OpenAI also received $350M in CoreWeave equity. Later expansions pushed the companies' announced contract value to roughly $22.4B.

OpenAI helps create CoreWeave revenue while simultaneously holding an economic interest in CoreWeave's value.

Nvidia is also one of CoreWeave's most important strategic partners.

A 2025 SEC filing shows Nvidia agreed to purchase residual unsold CoreWeave capacity through April 2032 under an agreement with an initial value of $6.3B.

CoreWeave fills data centers with Nvidia equipment.

Nvidia owns CoreWeave equity.

If CoreWeave cannot sell certain contracted capacity elsewhere, Nvidia can become the buyer.

That is a very different risk structure from a simple arms-length chip sale.

Equity is only part of the story.

AI infrastructure requires enormous amounts of debt.

Data centers need land, power, cooling, electrical infrastructure, chips and networking equipment years before their customers generate enough cash to pay for everything directly.

That has produced a growing market for special-purpose vehicles, private credit, long-term leases and residual-value guarantees.

The Financial Times recently reported that Big Tech companies have used guarantees to support as much as $300B of debt exposure tied to AI data centers and chips. The arrangements can allow infrastructure to sit outside the conventional balance sheet of the company ultimately supporting its economics.

One of the most striking examples involves SB Energy, Nvidia and OpenAI.

Nvidia announced in August that it would provide credit support for the land, power and shell infrastructure at SB Energy's massive PORTS-Pike project in Ohio. Nvidia is also investing $1.5B in SB Energy. OpenAI will be the customer under a 20-year lease, and the facility will use Nvidia's full-stack compute architecture.

The relationship looks like this:

Financiers provide capital
↓
SB Energy builds the data center
↓
Nvidia provides equipment + credit support + equity
↓
OpenAI signs the lease
↓
OpenAI's compute demand supports the project debt

The FT reports that Nvidia-backed structures have included residual-value guarantees covering up to 25% of some transactions, while Broadcom has provided roughly $29B of guarantees connected with Anthropic-related chip financing.

Google's Anthropic buildout provides another view of the machinery. The FT reported that a first roughly $35B tranche of TPU hardware was placed into a special-purpose vehicle financed with outside debt. Broadcom reportedly provided residual-value support for around $30B of that financing.

Private capital buys the hardware.

Anthropic leases the hardware.

Google and Broadcom provide pieces of the commercial and financial architecture.

The model makes enormous infrastructure spending possible without requiring every dollar of capex to sit directly on the customer's balance sheet.

It is easy to look at these arrows and reach for the most dramatic word available.

The underlying economics call for more precision.

Harvey has paying law firms.

Legora has paying legal departments.

Anthropic sells real inference.

AWS provides real compute.

Nvidia ships physical hardware.

Data centers consume real electricity and provide real capacity.

There is also nothing inherently unusual about several companies recognizing revenue from the same final pool of consumer spending. Every industrial supply chain works this way.

The more useful vocabulary is:

Revenue stacking: one customer's expenditure produces revenue at several stages of a supply chain.

Vendor financing: a supplier helps finance a customer that purchases its products.

Strategic cross-ownership: counterparties become investors in one another while maintaining commercial relationships.

Capacity backstops: a supplier, customer, or partner absorbs some risk if infrastructure cannot be placed with outside buyers.

Valuation reflexivity: higher valuations improve access to capital, which funds greater spending, which generates revenue elsewhere in the ecosystem and can help support additional valuations.

Those mechanisms can support a healthy buildout when the end demand is durable.

History also gives us a reason to pay attention.

During the telecom boom around 2000, capacity providers sometimes entered reciprocal fiber transactions. The SEC later found that Qwest used some capacity swaps to inflate reported revenue, with companies effectively buying capacity from one another while Qwest recognized revenue upfront and capitalized the corresponding purchase. Those transactions involved accounting practices and conduct materially different from today's disclosed AI partnerships, but the episode shows why investors care about the distinction between gross transactions inside an ecosystem and outside economic demand.

AI does not need to repeat that history for the lesson to matter.

The core issue is easier to see if you divide the money entering the AI ecosystem into two categories.

This is money coming from customers who want the technology because it provides economic utility:

These dollars originate outside the AI financing loop.

Then there is the capital circulating inside the stack:

The ecosystem can become much larger than the outside revenue currently entering it because capital markets are funding infrastructure for future demand.

That is common in periods of rapid infrastructure buildout.

The risk comes from correlation.

Anthropic and OpenAI reportedly account for more than half of roughly $2T in backlog across several major cloud providers, according to reporting cited by Reuters.

A huge cloud backlog can therefore look diversified across Amazon, Microsoft, Google and other infrastructure providers while a significant portion of the underlying economic demand traces back to a very small number of frontier labs.

The same concentration continues farther down the stack.

Sona Asset Management recently mapped 255 public companies connected to the AI buildout through investment, purchasing and financing relationships. Axios described the result as an ecosystem becoming increasingly interconnected financially.

That interconnectedness is powerful during expansion.

It also means a demand shock would travel.

Consider a much less dramatic scenario than an AI crash.

Enterprise AI spending simply grows slower than expected.

Application companies would see lower usage growth.

That reduces incremental demand for inference.

Model providers require less incremental compute than their infrastructure plans assumed.

Cloud providers receive fewer new commitments.

Some data-center capacity becomes harder to place.

Residual-value guarantees begin to matter more.

Lenders demand higher yields.

The next infrastructure project becomes more expensive.

Funding rounds get harder.

Companies have less capital available to sign the next generation of giant compute contracts.

Then the financial loop begins working in reverse.

There are already small signals that capital markets are becoming more selective. Reuters reported this week that investors are demanding wider spreads on AI-related corporate bonds as issuance increases, with concern centered on the scale and unpredictability of future infrastructure borrowing.

SB Energy also slowed its planned IPO amid scrutiny of its financing requirements and concentration around major AI customers. The company will require vast amounts of project-level capital to fulfill its data-center plans.

None of this establishes that AI demand is collapsing.

It shows where stress would appear first.

For early-stage investors, the lesson goes deeper than deciding whether AI is overvalued.

A startup can show explosive ARR growth while sitting on top of expensive variable inference.

That makes gross margin trajectory increasingly important. Harvey's experience is useful because a surge in product usage was simultaneously a sign of product-market fit and a major economic problem until the company changed how it delivered inference.

Revenue quality also matters.

ARR, annualized revenue run rate, contracted backlog, and recognized revenue are different metrics.

Cognition's $900M figure is an annualized run rate. Lovable's $500M figure is also annualized. Glean labels its $300M figure ARR, although part of its business uses consumption pricing. Comparing those numbers as though they represent identical contractual economics can produce bad underwriting.

Then there is supplier concentration.

A company whose product economics depend heavily on one model provider can see its margins change when model pricing changes. A startup that can route between models, fine-tune smaller models, or push workloads toward open-weight infrastructure has more control over its cost structure.

Finally, ask one question that rarely appears on a pitch deck:

Then follow it.

Who pays the startup?

Who does the startup pay?

Who finances those suppliers?

Do any of those suppliers own equity in the company?

Are contractual commitments being supported by guarantees from another participant in the same ecosystem?

The answers tell you considerably more than the headline ARR number.

AI is creating some of the fastest-growing companies venture capital has ever seen.

Harvey crossing $400M ARR within four years is extraordinary. Legora doubling from $100M to $200M in less than six months is extraordinary. Cognition adding hundreds of millions of annualized revenue between financing rounds is extraordinary. Anthropic moving from roughly $9B in run-rate revenue at the end of 2025 to more than $65B by July 2026 is on another scale entirely.

Those numbers deserve attention.

So does the structure underneath them.

AI has become an enormous interconnected capital system where enterprise budgets, venture funding, corporate balance sheets, private credit and infrastructure spending increasingly feed into one another.

The durability of the boom will ultimately depend on the amount of economic value reaching the customer at the edge of that system.

If an enterprise can spend $1 on AI and reliably create $2, $5 or $10 of value, the infrastructure behind it can support an enormous market.

If that return fails to materialize broadly enough, the industry has built a very large financial structure against demand that still has to arrive.

The AI boom increasingly operates as a financial flywheel.

Customer spending becomes application revenue.

Application usage becomes model revenue.

Model demand becomes cloud backlog.

Cloud demand becomes chip and data-center revenue.

Capital then flows back through the same network in the form of strategic investments, financing facilities, equity stakes and guarantees.

Every layer can represent genuine economic activity.

For investors, the number that matters most sits underneath all of them:

That is the number we should be watching.

Interested in investing alongside us?

At Dhow, we’re watching where the market is getting crowded and where it’s still being mispriced. AI is pulling record amounts of capital into venture, but the broader private market is still sorting through stale 2021 valuations, weak liquidity, and a widening gap between the companies everyone wants and everyone else. For early-stage investors, that makes entry price, ownership, dilution, and sourcing matter more than ever. The opportunity is still there, but the easy money is gone. Join the movement, share this with a friend (or two).

Blu Dot used Roku Ads Manager to drive incredible results for its furniture sales event. Its strategy hinged on custom audiences and retargeting, where intent was strongest.

“Roku has been a top performer,” said Blu Dot’s Claire Folkestad. “We have seen…CPMs lower than any other CTV partner we've worked with.”

Learn More

If you enjoyed this, you’d be doing us a HUGE FAVOR by just clicking the ad above & below - we appreciate it!

It’s your always-on revenue engine: agents and workflows build pipeline, chase every buying signal, and move deals forward alongside your team.

Teams like Parallel, Turbopuffer, and Wordsmith build on Attio. Are you one of them?

Try Attio now

Sources

This newsletter is for informational and educational purposes only and does not constitute investment advice, an offer to sell, or a solicitation to purchase any security. Private-market investments involve substantial risk, including loss of principal and illiquidity. Reported private-company revenue, ARR, run-rate, valuation and financing figures may be company-reported or sourced from third-party reporting and can use different methodologies.