Is Uber's Share of Ride Revenue Unfair?

It is easy enough to get an argument about the cost of using a marketplace such as the Apple App Store, Etsy or Uber. When Apple takes 15 percent to 30 percent of a sale, it can seem unfair to the seller. Uber’s take, said to be up to 50 percent in some cases, can seem usurious. 

But distribution, the roles in a value chain that move products or services from manufacturer to buyer, have a definite cost. 

And for some of us, a market maker or marketplace represents the cost of distribution. Seen that way, perhaps 30 percent is not unreasonable. 

Market makers, platforms and distributors all move goods between buyers and sellers. 

Function	Traditional Distributor	Marketplace / Market Maker (Uber, eBay, Etsy)
Takes ownership of product?	Usually yes	Usually no
Holds inventory?	Yes	No
Warehousing/logistics?	Major function	Usually little or none
Sets resale price?	Often yes	Usually seller sets price
Bears inventory risk?	Yes	No
Main asset	Physical network	Digital network
Revenue model	Gross margin on resale	Commission ("take rate")
Value provided	Physical distribution	Matching buyers and sellers
Scalability	Limited by physical assets	Highly scalable

A traditional distributor such as Sysco or McKesson buys products, stores them, transports them, and resells them.

A platform such as Uber, eBay, or Etsy generally does not own the underlying goods or services. Instead it creates a market, establishes trust, handles payments, provides discovery, and charges a fee.

The point is that platforms, market makers and distribution networks provide a similar function. 

From a value-chain perspective, both perform an intermediation function:

• "How do products get from factory to customer?"

• "How do buyers find sellers and transact safely?"

Both reduce search costs, transaction costs, and coordination costs.

The major innovation of digital marketplaces is that they perform many distribution functions without taking inventory ownership.

One could argue that Uber is a "virtual distributor" of transportation services, while eBay and Etsy are "virtual distributors" of goods.

So distribution represents a necessary part of any retail value chain. 

Research from the Reserve Bank of Australia found that for retail goods, roughly half of the final retail price reflects wholesale and retail distribution margins and costs. 

But distributor profits themselves were less than 10 percent of final sale price.

Final Retail Price = $100	Share
Manufacturing cost	$50
Wholesale distribution costs and margin	$15
Retail costs and margin	$35
Final consumer price	$100

Industry	Distributor Margin
Electronics	3%–10%
FMCG/Grocery	3%–10%
Industrial products	10%–20%
Medical products	20%–30%+
Apparel	15%–30%
source: Unanswered

Digital marketplaces usually charge commissions ("take rates") rather than earning resale margins.

Platform	Approximate Take Rate
Etsy	~6.5% transaction fee plus payment and advertising fees; effective cost often 10%-20%+ for sellers (Reddit)
eBay	Often around 10%-15% depending on category (Business Insider)
Uber	Company reports roughly 20%+ take rates, though some external studies estimate substantially higher in certain markets (Business Insider)

Looked at that way, take rates or distribution costs are quite similar. 

Role	Typical Share of Final Transaction Value
Physical distributor profit	3%-30%
Marketplace take rate	10%-30%
Combined wholesale + retail distribution system	Often 30%-50%+
Uber/eBay/Etsy platform fee	Often 10%-30% (50% in some cases for Uber)

Industrial Products	Digital Era Products
Scarcity = moving products	Scarcity = matching participants
Value = logistics	Value = network effects
Advantage = warehouses	Advantage = users and data

In economic terms, platforms replace physical intermediation with information intermediation.

So we move from:

Manufacturer → Distributor → Retailer → Consumer

to:

Producer → Marketplace → Consumer. 

As a result, a 10 percent to 20 percent marketplace fee can sometimes replace a traditional channel structure that consumed 30 percent to 50 percent of the final selling price.

source: Len Sherman 

In other words, a 30-percent take rate for sellers on a platform or marketplace might seem out of line, but might not actually be usurious. 

Compulsory License and AI Copyright

Language models are creating new terrain for copyright law.

Personally, I favor as much freedom as we can possibly endure. On the other hand, practical economic realities are likely to dictate outcomes that balance payments to rights owners and encouragement of innovation by AI firms. 

History strongly suggests that pattern will emerge. 

In the past, rights holders usually wanted control (veto power over new uses). Courts and Congress usually compromise by substituting compensation (compulsory licenses, collective societies, levies) for control. 

Radio stations, cable companies, physical media distributors and streaming platforms captured enormous value from others' content, then eventually reached licensing accommodations. 

Player piano rolls in the late 1800s reproduced sheet music mechanically, and publishers argued this was infringement. 

The Supreme Court ruled in White-Smith Music v. Apollo (1908) that piano rolls were not copies because copyright covered notation readable by humans, not mechanical reproductions.

Congress then passed the Copyright Act of 1909, which created the  compulsory mechanical license, under which anyone could record a song already recorded by someone else, provided they paid a statutory royalty.

Radio created a new problem:

• Composers and publishers got performance royalties through ASCAP (founded 1914), which negotiated blanket licenses with broadcasters.

• Performers and record labels got nothing. The law treated a broadcast performance as fundamentally different from a mechanical reproduction, so the actual recording artists received no royalties when their records were played on air.

Film created other issues. Sound film after 1927 meant a single film now embodied multiple copyrights simultaneously:

• the screenplay

• the musical score

• the recorded performances

• the film itself. 

The question of who owned what in a collaborative industrial production led to the work-for-hire doctrine. Studios owned everything, employees and contractors owned nothing.

International distribution exposed the territorial nature of copyright immediately. The Berne Convention (1886, continuously expanded) became the framework for international harmonization.

Magnetic tape, then cassettes, then VHS created successive waves of the same basic problem: cheap, accessible reproduction technology. 

The 1971 Sound Recording Amendment was the first law to give sound recordings their own copyright protection (previously, only the underlying composition was protected).

Sony v. Universal City Studios (the "Betamax case," 1984) is one of the most consequential copyright decisions ever. Hollywood sued Sony for making VCRs, arguing Sony was liable for the infringement its customers committed. 

The Supreme Court ruled 5-4 that:

• Time-shifting (recording TV to watch later) was fair use

• Selling a device with substantial non-infringing uses didn't make the manufacturer liable.

The Betamax decision shaped technology law for decades. It's why VCRs, DVRs and the consumer electronics industry could exist without needing Hollywood's permission. 

The Audio Home Recording Act (1992) eventually addressed digital audio tape (DAT) by requiring copy-protection technology in DAT recorders and adding a levy on blank digital media, a compromise that established the principle that hardware makers could be taxed to compensate rights holders.

Cable TV initially retransmitted distant television broadcast signals without compensation. The Supreme Court ruled twice (1968, 1974) that cable retransmission wasn't infringement because cable companies weren't "performing" the works. 

But the Copyright Act of 1976 created a compulsory license for cable retransmission.

This compulsory-license-as-compromise model recurs throughout copyright's adaptation to new media.

Xerography created frictionless reproduction of printed text.

The 1976 Act addressed this partly, and the Copyright Clearance Center was founded in 1978 as a collective licensing organization, allowing libraries and businesses to pay blanket fees for photocopying rights.

CDs introduced a paradox: perfect digital copies. This era produced the first serious deployment of Digital Rights Management concepts. It also sharpened the debate about the first sale doctrine — you can resell a CD you bought, but can you resell a digital file? 

The DMCA (1998) was the legislative response. 

MP3 and the iPod era forced the unbundling of the album into individual songs, collapsing per-unit revenue and forcing the industry toward licensing models.

Streaming is philosophically interesting because it doesn't fit the reproduction model at all — you're not copying anything, you're accessing a performance. This is actually closer to the original concept of performance rights than anything since radio.

Spotify, Netflix, and their successors forced the creation of entirely new licensing frameworks:

mechanical licenses for on-demand streaming, negotiated rates between platforms and labels, windowing strategies for film. 

The Music Modernization Act (2018) was the most significant music copyright reform in decades, largely cleaning up the licensing mess streaming had created.

Crucially, streaming finally gave performers (not just composers) digital performance royalties. 

AI companies likely will eventually fit this compulsory license pattern almost perfectly.

The internet produced many precedents that might shape AI copyright:

• The Digital Millennium Copyright Act

• The notice-and-takedown system (rights holders can demand platforms remove infringing content, and platforms that comply get "safe harbor" protection from liability

• Anti-circumvention rules making it illegal to bypass digital rights management (DRM) technology

• ISP liability limits** — shielding internet service providers from liability for what users transmit, provided they act on takedown notices

• The No Electronic Theft Act (1997) closed a loophole where non-commercial infringement wasn't clearly criminal. Before it, you had to be *selling* pirated content to face criminal liability

• The Sonny Bono Copyright Term Extension Act (1998) extended copyright terms by 20 years

• Perfect 10 v. Amazon/Google (2007) established that search engine thumbnail images and inline linking could qualify as fair use

• Authors Guild v. Google validated Google's mass scanning of copyrighted books as fair use

• Viacom v. YouTube (2012) confirmed that platforms don't lose protection simply because infringement is general knowledge

• The Napster and Grokster cases (2001, 2005) established "contributory infringement" and "inducement" doctrines: you can be liable not just for infringing yourself, but for building a platform *designed* to encourage infringement. 

Beyond law, the internet forced copyright to adapt through raw economic pressure:

• Licensing at scale became essential

• Creative Commons (2001) created a voluntary licensing system letting creators specify in advance what reuse they permit

• Terms of service became a shadow copyright system

• Geoblocking and regional licensing became standard business practice as companies realized the internet didn't respect territorial copyright boundaries that the entire global licensing system was built around.

The historical pattern suggests the outcome will likely be some combination of new licensing frameworks, platform liability rules, and compensation mechanisms rather than either "AI training is fully free" or "AI training is categorically infringing." 

The internet precedents suggest the law will bend toward enabling the technology while extracting some structural protections for creators.

My own thinking leans towards permissiveness where it comes to the use of content. As for the argument that language models unfairly infringe copyrights because they “read” or “ingest” content, my argument would be that this non-infringing work is what humans do routinely, and it is not copyright infringement. 

When a person reads a novel, watches a film, or listens to music, they:

• Absorb patterns, styles, vocabulary, narrative structures

• Develop taste and skill influenced by what they've consumed

• Produce new works that are clearly downstream of that consumption

• Do all of this without paying royalties or seeking permission

Nobody considers this a copyright violation, even when the influence is obvious. A novelist can write "in the style of Hemingway" after reading all his books, or a musician whose sound is clearly shaped by artists they grew up listening to.

AI models do that, but with some important differences, some will argue:

• Scale and speed

• Reproduction during training, which involves making copies of copyrighted content and storing them (transiently) on servers. Courts have historically treated that mechanical act of copying as significant, regardless of what happens downstream

• Memorization and regurgitation at scale that arguably collapses the distinction between learning from something and copying it

• Market substitution when an AI trained on a photographer's portfolio can produce images that replace demand for that photographer's work.

The enduring issues suggesting preserving freedom are:

• knowledge and style aren't ownable

• culture builds on culture

• the internet was built on the assumption that content could be read and indexed.

The case for creator protection rests on: 

• copyright exists precisely to ensure creators can sustain the work of creation

• the AI industry is extracting enormous commercial value from creative labor without compensation

• "humans do it too" ignores that humans don't build billion-dollar products directly monetizing others' uncredited work.

The resolution will require development of compensation structures that make the ecosystem fair to creators while not strangling a genuinely transformative technology.

In my view, compulsory license in some form.

SpaceX Acquisition of Cursor is About the Stack

SpaceX is acquiring Cursor’s parent company Anysphere for $60 billion, and the valuation might be more a matter of strategic value than traditional software as a service metrics.

In other words, critics will question the cost of the deal to SpaceX.

Cursor already had become the fastest-scaling enterprise SaaS company ever, growing from zero revenue to $2 billion annualized run rate in three years. 

If SpaceX acquires Cursor, the biggest advantages would be faster entry into AI coding, access to Cursor’s loyal developer base, and a stronger product layer for monetizing its compute infrastructure. 

For Cursor, competition from Claude Code seems key. But Cursor also gains from SpaceX compute infrastructure, as compute resources have been an issue in the past. 

source: MarketWatch 

It should therefore provide a better competitive position against Anthropic in enterprise AI.

Cursor is a fork of Microsoft’s Visual Studio Code, the most-widely-used code editor in the world, with AI capabilities integrated at every level of the development workflow. 

It autocompletes code, suggests changes across multiple files, runs tests, iterates on errors, and increasingly operates as an autonomous agent that can execute multi-step coding tasks with minimal human intervention. 

But Cursor has been dependent on third-party AI providers for the underlying intelligence. 

As those model developers increasingly launched competing coding products, Cursor risked being squeezed by companies that controlled both the models and the computing infrastructure. 

For SpaceX, the deal adds a fast-growing software business to an AI platform built around xAI and the company's Colossus data center infrastructure.

The deal would also help SpaceX compete more directly in AI agents and coding tools, where practical developer utility matters as much as raw model quality.

Buying Cursor does not automatically close the gap with the best frontier-model labs, though. 

SpaceX would still need to keep improving model quality, not just own a popular application layer. But the move leads to a more integrated approach (energy, compute infrastructure, models, apps).

Mergers, Joint Ventures or Investments as Routes to Controlling AI Model Costs

Just how artificial intelligence model providers might improve their economics is a key business model issue. 

A shift to inference operations also emphasizes the importance of reducing cost per token at scale. 

source: McKinsey 

Where software often has marginal costs close to zero, use of AI models seem to have costs that scale almost linearly with usage, so marginal costs are high. 

A few key business issues are clear enough.

A greater shift of cost towards fixed cost rather than variable cost seems necessary. For example, model creators could move in the direction of owning their compute infrastructure rather than renting cloud capacity.

The problems with high variable cost are clear:

• Unit economics

• Cash burn and need for capital injections

• Competitive pressure

• Capital allocation

Right now, model queries do not show software-style marginal cost trends, where marginal costs are close to zero. 

Every additional user or query drives proportional costs for GPUs, power, and data center capacity. 

Reports suggest inference alone consumes 50 percent of revenue in some cases. Some suggest the problem is worse, with at least some model providers spending more than 100 percent of revenue on compute services. 

By itself, that might not be existential, as model providers are in the early stages of growth, meaning incremental revenues would not be expected to cover the full costs of creating and operating the models at scale.  

But observers do worry about marginal costs that seemingly do not have software-style economics with near-zero marginal costs.

Cash burn and capital intensity also are issues. Rapid revenue growth is offset by even faster cost scaling, which necessitates high investor funding requirements, borrowing, equity raises. 

Pricing issues also are an issue. When model suppliers raise prices, introduce usage-based tiers, or limit free access, they risk customer churn or slower adoption, even as they address revenue issues. 

Strategic vulnerabilities also exist when model suppliers are dependent on external suppliers for crucial computing services (operating costs, ability to manage surges of demand). 

Capital allocation is an issue as well. One might argue that model builders should divert capital into compute infrastructure, but that is hugely expensive and detracts from the job of developing the next generation of models.

So the issue is how to fix the problem. Revenue growth with scale obviously helps, but doesn't solve the marginal cost issue. Efficiency improvements, owned infrastructure and pricing innovations all will play a role.

Smaller or distilled models, sparse activations, better architectures, prompt caching, batching, quantization, and routing to cheaper models for simpler tasks will happen.

Inference costs per token have dropped significantly in some cases, allowing gross compute margins to  improve.

source: Beth Kindig 

Stranded assets always are a problem, so higher GPU utilization rates help. So do custom silicon and algorithmic advances.

Owned infrastructure is a partial answer. Model builders and compute suppliers are investing heavily in their own data centers and chips (Anthropic's $50 billion commitments for custom U.S. facilities with partners like Fluidstack; OpenAI's Stargate).

Revenue models also are adjusting. Higher enterprise pricing, usage-based tiers, value-based pricing (charge relative to delivered value, not just tokens) and premium features or apps are introduced. 

Given all that, one logical historical precedent is for mergers and acquisitions that place model building and compute functions under a single ownership. On the other hand, antitrust regulations will probably tend to restrict options for some of the most-likely buyers (Alphabet, AWS, Microsoft, SpaceX, for example).

So other forms of cooperation are likely to develop. 

Expect partnerships, joint ventures, dedicated capacity deals, and partial ownership rather than full-scale mergers and acquisitions, which will face antitrust opposition. 

When feasible, model builders are creating their own compute infrastructure. OpenAI's Stargate project with Oracle and SoftBank (up to $500B, multi-gigawatt scale) provides an example. Investments in neocloud suppliers is another example. 

But that might be the exception to the rule. Anthropic, for example, has chosen to sign big supply deals rather than build its own facilities.

Partnerships for Tensor Processing Units, Trainium, and other accelerators reduce reliance on expensive third-party GPUs and improve efficiency also are growing. 

But full vertical integration might not be the immediate or mid-term path forward, partly for regulatory scrutiny reasons; partly for capital intensity reasons and partly for business diversity reasons. Both model builders and compute infra providers prefer a diversity of partners. 

So joint ventures and consortia that also have the advantage of off-balance-sheet implications will happen. 

In summary, tight strategic integration and partial ownership rather than blockbuster mergers are the main approaches. That avoids regulatory opposition and also is capital efficient. 

Solving AI Model Marginal Cost Issues

Profit margins arguably are the key business issue for frontier artificial intelligence model providers.

Where software businesses have tended to have low marginal costs, with high profit margins at scale, AI businesses tend to have no such advantage. 

Inference and training have heavily variable costs tied to usage and scale. In other words, cost per query scales mostly in linear fashion. 

That is quite unlike traditional software, with 80 percent to 90 percent gross margins at scale, since marginal cost is quite low. 

And that might have important ramifications for pricing models.

Traditional software can often support very high gross margins because the marginal cost of serving one more user is close to zero, while AI products can have materially higher variable costs because each additional query or token consumes compute.

That difference pushes AI businesses toward pricing and operating models that track usage, outcomes, or hybrid bundles rather than pure seat-based subscriptions.

For classic software, once the product is built, distributing another copy is cheap, so revenue can scale much faster than cost. 

In AI, each extra interaction may add inference, context, storage, and orchestration costs, so the economics can look more like a utility or metered service than a pure software license.

That does not mean AI cannot be highly profitable; it means profitability depends more on model efficiency, pricing design, and control over the inference stack.

The emerging pattern is hybrid monetization mixing flat rate charges and usage:

• subscriptions for baseline access

• usage-based credits

• metered throughput

• outcome-based pricing for heavier AI usage.

If additional usage carries significant marginal cost, sustainable AI business models usually combine four elements: 

• cost-aware product design (cost per interaction)

• pricing aligned to actual usage intensity (tiered pricing)

• infrastructure leverage (own rather than rent)

• AI feature packaging (proprietary data, workflow integration). 

A practical way to think about it is this: the business must make sure marginal revenue per interaction stays above marginal cost per interaction, with enough spread to fund sales, support, R&D, and model improvement. 

That often means charging more for complex tasks, throttling or downgrading expensive requests, or bundling some usage into plans while monetizing power users separately.

Issue	Why it matters	Possible solution
High per-use compute cost	Margins can compress as usage rises	Tiered pricing, credits, or metered billing tied to task intensity
Third-party model dependence	Vendor price changes and rate limits can hit gross margin	Own more of the inference stack, add model abstraction, or use open-source/smaller models where feasible
Weak price predictability	Customers dislike uncertain AI bills, slowing adoption	Pre-commits, usage caps, transparent dashboards, and clear credit buckets
Flat seat pricing mismatch	Power users can consume far more than average users	Hybrid subscription plus usage-based overages or action-based pricing
Hard-to-measure outcomes	Outcome pricing is difficult to verify and negotiate	Start with activity-based metrics, move to outcomes only where measurable and automated
Scaling costs faster than revenue	Growth can worsen economics instead of improving them	Cost monitoring, workload routing, batch inference, and feature-level margin analysis
Commoditization risk	Models alone are easy to copy	Build proprietary data, vertical specialization, and workflow integration

AI winners are more likely to be firms that control costs, can manage usage and price in a way that reflects both value delivered and the true marginal cost of serving that value.