David Weaver

Posted on Aug 10

Where Will Durable AI Competitive Advantage Accrue?

#ai #product #venture #supplychainofintelligence

The Supply Chain of Intelligence framework: The best structural answer to the question every investment committee is asking

I build at the application layer of AI, which means I have spent three years being told my company is a wrapper by people using a map that could not explain Cursor, Chegg, and NVIDIA at the same time.
Inside Y Combinator’s Mountain View offices, some partners began circulating a shared internal reference they simply called “SCoI” At a16z, partners on the infrastructure and growth teams started annotating every AI deal memo against a ten-layer and fifty-sub layer map that was being evaluated.
The framework they are talking is the Supply Chain of Intelligence, developed by Anand Arivukkarasu, a former Meta product leader, and published at supplychainofai.com. This essay lays out that framework and tests it against what the industry's most rigorous investors have published and what the market has actually done. Every quotation is verbatim from a cited source. Every figure carries its source and date. Where a number is reported rather than audited, the text says so.*

I. The most expensive open question in technology

In the first half of 2026, global venture investors deployed a record $510 billion, more than they invested in all of 2025, and by the second quarter more than seventy percent of it was going to AI companies, according to Crunchbase data published in July 2026. That is the largest concentrated capital allocation to a single technology thesis in the history of venture investing.

It is being deployed into a stack that the industry's most sophisticated observers have repeatedly described as having no obvious defenses. In January 2023, Andreessen Horowitz's Matt Bornstein, Guido Appenzeller, and Martin Casado published the essay that framed the era: "There don't appear, today, to be any systemic moats in generative AI." They went further: "So far, we've had a hard time finding structural defensibility anywhere in the stack, outside of traditional moats for incumbents." Eight months later, Sequoia's Sonya Huang and Pat Grady published their own correction of the industry's founding assumption, writing in "Generative AI: Act Two" that "the moats are in the customers, not the data," and conceding that the proprietary-data thesis they had championed a year earlier had not held: application-layer data "does not create an insurmountable moat."

So the state of play, stated plainly: half a trillion dollars a year is chasing an asset class whose leading underwriters have publicly documented their difficulty locating the moat. Either the capital is wrong, or the maps are.

I believe the maps are. The frameworks investors inherited, network effects, switching costs, scale economies, brand, were built to describe software businesses. AI is not a software business with a new feature. It is a new industrial input, and industrial inputs are analyzed with a different tool: the supply chain. That is the reframe at the center of Arivukkarasu's Supply Chain of Intelligence, and when you redraw the AI economy his way, ten layers running from physical resources to persistent memory, the confusion resolves. Durable advantage is not accruing to "AI companies." It is accruing to specific, nameable layers of the chain, and it is being destroyed, with equal precision, at others. The evidence for this is no longer theoretical. Three years of funding data, revenue trajectories, and corporate obituaries now trace the pattern clearly, and the sharpest investors have each independently discovered fragments of it. What follows is the whole map.

II. Why the old maps fail: intelligence is deflating

Begin with the fact that breaks every conventional analysis: the core commodity of this era is collapsing in price faster than any industrial input in economic history.

Guido Appenzeller of a16z documented the rate in November 2024 and gave it a name, LLMflation: "For an LLM of equivalent performance, the cost is decreasing by 10x every year." His data: when GPT-3 became publicly accessible in November 2021, achieving a benchmark score of MMLU 42 cost $60 per million tokens; by late 2024, an equivalent-performing open model cost $0.06 per million tokens. A thousandfold decline in three years. The trajectory has continued. OpenAI's published API prices fell from $30 per million input tokens at GPT-4's launch in March 2023 to $2.50 for GPT-4o by May 2024, and Google priced Gemini 1.5 Flash at $0.075. By August 2025, Sarah Wang and Casado were writing that "inference costs have dropped anywhere from 10x to 100x+ in the last 18 months."

Deflation of this violence has a specific economic consequence: anything whose value consists mainly of reselling the deflating input gets repriced toward zero. This is not a metaphor. It is the observable mechanism behind every major AI casualty of the last three years, and it is the first law of the framework.

But deflation alone would predict that no one makes money, and that prediction fails spectacularly. NVIDIA reported a record $75.2 billion in data center revenue in a single quarter in May 2026, up 92 percent year over year, at a 74.9 percent gross margin. Cursor's maker Anysphere went from a reported $100 million to $2 billion in annualized revenue in roughly thirteen months. Harvey, the legal AI company, raised at an $11 billion valuation in March 2026 with reported revenue tripling. Menlo Ventures' enterprise survey found companies spent $37 billion on generative AI in 2025, up 3.2 times year over year, with the application layer taking 51 percent of it and startups, not incumbents, capturing 63 percent of that application revenue.

Enormous value is being captured. It is just not being captured where the phrase "AI company" points. To see where it actually pools, you need the chain.

III. The Supply Chain of Intelligence

The familiar three-tier stack diagram, hardware, models, applications, has become marketing furniture. It describes how intelligence is assembled. It says almost nothing about where economic power settles, which is the only thing an investor or founder actually needs a map for. A supply chain view fixes this, because supply chains are drawn to answer precisely that question.

Every industrial economy organizes around a supply chain, and profits never distribute evenly along it. They pool at the scarce links. Oil economics pooled at refining and distribution, not at the derrick. PC economics pooled at the processor and the operating system, not at the assembler. The generative AI economy is an industrial economy for intelligence, and it has a supply chain like any other. The framework maps it in ten layers:

Layer minus 1, Resources: energy, thermal and water capacity, fabrication, critical materials, skilled trades. Layer 0, Infrastructure: silicon, data centers, interconnect, compute clouds, edge devices. Layer 1, Data: public, proprietary, behavioral, outcome, and synthetic data. Layer 2, Models: foundation, specialized, embedding, routing, reasoning. Layer 3, Gatekeeping: compliance, quality, safety, editorial judgment, distribution control. Layer 4, Access: APIs, agent protocols, governance, real-time infrastructure, agent identity. Layer 5, Execution: domain execution, decision frameworks, retrieval workflows, operating playbooks, interactional skill. Layer 6, Orchestration: agent loops, human-in-the-loop design, task decomposition, context management, runtime assurance. Layer 7, Surface: the interfaces where users meet intelligence, conversational, visual, embedded, transactional, ambient. Layer 8, Memory: session memory, user profiles, network-level learning, institutional knowledge, learned world models.

The framework's central claim is that these layers are not equally defensible, and that four structural laws govern how value moves between them. The laws are not aspirations. Each one now has a body count and a set of compounding winners, and each one, as I will show, has been independently discovered in fragments by a16z, Sequoia, Y Combinator, Bessemer, and Menlo, without any of them assembling the full chain the way this framework does.

IV. Law One: intelligence commoditizes downward

If your product's value depends on generic model capability, the layer below you will absorb you. Wrappers become features.

The evidence here is the most complete, because this law has been running the longest.

Jasper raised $125 million at a $1.5 billion valuation in October 2022, the emblematic company of the first generative AI wave, selling AI copywriting on top of GPT-3. ChatGPT launched five weeks later and gave away the underlying capability. Within a year, per reporting from The Information, Jasper had cut its revenue forecast by at least thirty percent, marked down its own internal valuation by twenty percent, laid off staff, and replaced its CEO. Nothing about Jasper's execution failed. Its layer failed. The company owned a surface, Layer 7, on top of rented intelligence, and the intelligence commoditized underneath it.

Chegg is the public-market version, and its own executives wrote the diagnosis. The company lost more than half a million subscribers after ChatGPT's launch; by late 2025 it had cut 45 percent of its workforce, with CNBC reporting the company blamed "new realities of AI," and its stock sat 99 percent below its 2021 peak, roughly $14.5 billion in market value erased. Former CEO Dan Rosensweig described himself as the "poster child" of the AI shock, per Fortune. Stack Overflow tells the same story without a stock ticker: monthly questions collapsed from a peak above 200,000 to under 50,000 by late 2025, a level the site last saw around 2008, as its own developer survey showed 84 percent of developers now using AI tools. And the compression is not limited to the weak. Cognition's Devin, the most hyped AI software engineer of 2024, launched at roughly $500 per month; by April 2025 it was repriced at $20 per month, a 96 percent cut in entry price, as platforms bundled competing capability for free.

The pattern generalizes. In 2020, Casado and Bornstein had already observed in "The New Business of AI" that "the moats for AI companies appear to be shallower than many expected. AI may largely be a pass-through, from a defensibility standpoint, to the underlying product and data." That sentence, written before ChatGPT existed, reads today as a prophecy Jasper and Chegg fulfilled. A pass-through business earns pass-through economics.

One refinement matters, because the industry has overcorrected twice on this point. "Wrapper" was first a category, then an insult, and is now, in some corners, a badge of honor. Casado himself has argued on the a16z podcast in 2025 that the wrapper insult misfires, that nobody calls software built on the cloud a "cloud wrapper," and that models are best understood as a new infrastructure layer with "tremendous opportunity to add value above the stack." Andrew Chen made the same turn in his February 2025 essay "Revenge of the GPT Wrappers," noting that "state-of-the-art models only seem to be able to stay ~6 months ahead of their open source cousins" and that defensibility therefore migrates to the classic application-layer assets, network effects chief among them. Both are correct, and neither contradicts the law. The law does not say building on models is fatal. It says depending on nothing but the model is fatal. The question was never whether you wrap. It is what else you own. Which is precisely where the next three laws take over.

V. Law Two: value accrues at bottlenecks

Durable value sits at whichever layer is scarce: proprietary data, workflow control, verification, distribution, compliance, trust.

The supply chain frame earns its keep here, because it predicts something the software frame cannot: that the biggest winner of the AI era so far would be a component maker. NVIDIA sits at the tightest bottleneck in the chain, Layer 0 silicon, and its results read like a tax receipt on the entire industry: $75.2 billion of data center revenue in one quarter at margins near 75 percent. The a16z platform essay saw this early: "infrastructure vendors are likely the biggest winners in this market so far, capturing the majority of dollars flowing through the stack," and, in the essay's most quietly radical observation, that the companies creating the most value, meaning those training the models and building the applications, "haven't captured most of it." Value creation and value capture separated cleanly along the chain, exactly as supply chain economics predicts and software economics does not.

But bottlenecks are not only physical. Consider what the compounding application companies actually own. Harvey did not win legal AI by having a better interface to a model; it accumulated proprietary legal work product, embedded itself in the compliance-saturated workflows of more than 1,300 organizations and a reported 100,000 lawyers, and built institutional memory inside the firms it serves. Its valuation went from $3 billion to a reported $11 billion inside thirteen months, with a further raise at $15.5 billion reported to be in talks by The Information in August 2026, on revenue reported to have grown from roughly $190 million to over $300 million annualized. OpenEvidence, which reached a reported $12 billion valuation in January 2026, is used daily, per Crunchbase reporting, by more than 40 percent of American physicians: its bottleneck is not the model, it is verified medical trust and distribution into clinical workflow, Layers 3 and 5, at a scale no competitor can quickly replicate. Abridge doubled its valuation to $5.3 billion in four months on the strength of deployment inside more than 150 of the largest US health systems, a distribution and integration bottleneck, not a model advantage.

Menlo's 2025 enterprise data quantifies the shift in the aggregate: of $37 billion in enterprise generative AI spend, the application layer took $19 billion, and 76 percent of AI use cases are now purchased rather than built internally, with buyers converting at 47 percent against SaaS's historical 25 percent. Money is flowing to whoever resolves a scarcity, and in the enterprise the scarcity is no longer intelligence. It is trusted, integrated, verified execution.

VI. Law Three: surface captures attention; chain captures power

A beautiful interface gets you users. Owning a deeper layer keeps them.

Sequoia's "Act Two" essay supplied the defining statistic of surface-only economics: generative AI applications showed a median DAU over MAU ratio of 14 percent, against 51 percent for WhatsApp. Their verdict: "Hype and flash are giving way to real value and whole product experiences," and the honest admission that "generative AI's biggest problem is not finding use cases or demand or distribution, it is proving value." Casado made the same observation from the other coast, arguing on the a16z podcast that these models are so capable that wrapping one solves the customer acquisition problem instantly, and solves the retention problem not at all.

Cursor is the case that seems to refute this law and instead proves it. On the surface, literally, Cursor is an interface to other companies' models, the thing the wrapper critique said could never endure. Its reported trajectory: $100 million annualized revenue in January 2025, $500 million by June, $1 billion by November, $2 billion by February 2026, with a $29.3 billion valuation round reported by CNBC in November 2025 and subsequent reporting of talks near $50 billion. The fastest revenue scaling in the history of business software did not come from owning a frontier model. It came from owning everything between the model and the work: the integration into the development environment where code actually happens, the orchestration of multi-step coding agents, and the accumulating memory of each codebase and team. In the framework's terms, Cursor owns Layer 4 access, Layer 6 orchestration, and Layer 8 memory, and rents Layer 2 from whoever is best this quarter. The a16z 2025 CIO survey found one CTO reporting that nearly 90 percent of his company's code was being generated through tools like Cursor and Claude Code, up from 10 to 15 percent a year earlier, and the same survey's authors observed that "increasingly complex AI workflows are driving higher switching costs." The surface is what you see. The chain position is what you cannot leave.

Sequoia's 2024 essay "Generative AI's Act o1" states the deeper-layer thesis in language any founder can operationalize: "Application layer AI companies are not just UIs on top of a foundation model. Far from it." And: "The messy real world requires significant domain and application-specific reasoning that cannot efficiently be encoded in a general model." Sarah Wang locates the exact coordinates of the defensible zone: "where you have complex workflows and a ton of customer data where deep integrations actually are necessary to get that last mile value for the customer. This is where the specialized AI apps are sort of crushing any either foundation model layer or otherwise company in the market."

Sierra, Bret Taylor's customer-service agent company, adds the final piece: it charges, per TechCrunch, "for completed work rather than flat subscription fees," reaching a reported $100 million in annualized revenue within 21 months and a $10 billion valuation. Outcome pricing is only possible for a company that controls execution and can stand behind results, which is to say, a company operating at Layers 5 through 8. You cannot price on outcomes from the surface. The chain is what makes the business model available at all.

VII. Law Four: generation and verification must separate

Wherever output carries fiduciary, regulatory, safety, or reputational weight, the entity that generates cannot be the entity that certifies.

This is the least priced-in of the four laws, and it should be read partly as observation and partly as prediction, labeled as such. The observation: as agents move from drafting text to executing work, enterprises are already reorganizing around verification and control. The Futurum Group's February 2026 survey of 830 IT decision-makers found 41 percent of organizations actively consolidating applications, best-of-breed procurement falling to 20.7 percent, and the streamlining of AI agents ranked as a leading motivation; one analyst noted that organizations stitching together ten to fifteen point solutions face an integration tax that makes enterprise-wide AI strategy nearly impossible. Consolidation is a demand for governable, verifiable chains, not clever endpoints. YC's current Requests for Startups includes AI-native compliance infrastructure on exactly this logic, with partner Daivik Goel noting that most compliance work is "monitoring regulatory changes, flagging anomalies, generating reports," which is to say, verification work.

The prediction, which follows from the law rather than from data: the largest un-built companies of the agent era are verification layers, the entities that certify what agents did, audit what they touched, and carry the liability the generator cannot. When an AI writes the contract, someone must be paid to be the one who did not write it and can therefore judge it. Every mature industry ends up with this separation: auditors apart from accountants, ratings agencies apart from issuers, inspectors apart from builders. Intelligence will be no different, and Layer 3 is where that value will pool. This is a judgment, not yet a measured result, and it should be tested the way all such claims should be, against the next three years of company formation.

VIII. The convergence: five firms, one map

Here is the observation that convinced me this framework describes something real rather than something merely tidy. The major firms have each, independently and in their own vocabulary, discovered a piece of the chain, and the pieces do not overlap. They tile.

a16z discovered the deflation and the infrastructure capture: no systemic moats in the middle of the stack, infrastructure winning, inference deflating tenfold yearly, and by 2025, in Wang and Casado's words, value accruing to complex workflows, deep integration, and switching costs. That is Layers 0, 2, 4, and 5 of the map. Sequoia discovered the surface trap and the execution layer: retention failure at the interface, "moats are in the customers," workflows and user networks as the durable assets, selling work rather than software. That is Layers 5, 6, and 7. Bessemer's 2025 State of AI report discovered the top of the map in a single sentence: "Memory and context are the new moats. The most defensible products will remember, adapt, and personalize." That is Layer 8, stated almost verbatim. And their vertical AI thesis, that "defensibility stems from domain expertise: integrations, data moats, and multimodal interfaces built for vertical-specific needs," is Layers 1 and 5. Menlo quantified the flow of money up the chain: application layer at 51 percent of enterprise spend, startups capturing 63 percent of it. Y Combinator discovered the demand side: 46 percent of its Spring 2025 batch building AI agents per PitchBook, its partners arguing publicly that vertical AI companies could be ten times larger than the SaaS incumbents they replace, because, as a16z's Alex Rampell frames the prize, worldwide SaaS is roughly a $300 billion market while US labor alone is $13 trillion.

The chain also gives classical strategy theory its missing coordinates. Hamilton Helmer's Seven Powers, the most rigorous general theory of competitive advantage of the last decade, names the forms power takes: cornered resources, process power, switching costs, counter-positioning, scale and network economies, branding. What it does not tell an AI investor is where in this particular economy each power can physically live. The layer map supplies the address. A cornered resource in AI is Layer 1 proprietary data or Layer 0 capacity. Process power is Layer 5, the unglamorous encoded judgment that makes agents reliable against real-world edge cases. Switching costs live at Layers 4 and 8. Counter-positioning is what a Layer 5 outcome-priced company does to a seat-priced incumbent. The Seven Powers say what advantage is; the chain says where it can be built.

Each firm is describing the elephant from where it stands. The supply chain is the elephant. And once you see the whole animal, the investment question stops being "is this an AI company?" or even "is this a wrapper?" and becomes the only question that has ever mattered in industrial economics: which links of the chain does this company own, and how scarce are they?

IX. The Defensible Triangle, and the map in one table

Across every compounding case in the evidence, the same three-layer combination recurs, frequently enough that the framework gives it a name: the Defensible Triangle. Proprietary data (Layer 1) that competitors cannot fetch. Deep execution (Layer 5), the encoded playbooks, domain judgment, and workflow control that turn model capability into finished work. Compounding memory (Layer 8), the accumulated institutional and network learning that makes the product better for the ten-thousandth customer than the first. A company holding all three corners is effectively unrentable: the model beneath it can commoditize completely, and the business barely notices. The alternative survival strategy is absolute ownership of a single layer, the NVIDIA position, but that path is mostly closed to new entrants. For founders and investors at the application layer, the Triangle is the map.

Company	Layers owned	Structural position	Outcome (reported, with dates)
Jasper	L7 surface only	Rented intelligence, owned interface	$1.5B valuation Oct 2022; forecast cut ~30%, internal markdown, CEO change within a year of ChatGPT
Chegg	L7 + pre-AI content	Disintermediated by the model layer	Stock down 99% from 2021 peak; 45% of workforce cut in 2025
Devin (Cognition)	L7 on rented L2	Priced against bundled free capability	Entry price cut from ~$500 to $20 per month in one year
Cursor	L4 + L6 + L8	Chain position under a rented model	Reported $100M to $2B ARR in ~13 months; $29.3B round Nov 2025
Harvey	L1 + L3 + L5 + L8	Full Defensible Triangle plus compliance	Reported $3B to $11B valuation in 13 months; ~$300M ARR by Aug 2026
Sierra	L5 + L7d + L8	Outcome-priced execution	Reported $100M ARR in 21 months; $10B valuation
OpenEvidence	L1 + L3 + L5	Verified trust and clinical distribution	Reported $12B valuation; used daily by 40%+ of US physicians
NVIDIA	L0 absolute	Single-layer bottleneck ownership	$75.2B data center revenue in one quarter, ~75% gross margin

X. Where durable advantage will not accrue

A framework that only explains winners is astrology. The chain also identifies where advantage is structurally unavailable, and these are the positions absorbing the most capital today.

It will not accrue to undifferentiated model capability. The frontier labs are extraordinary businesses by revenue velocity, OpenAI reaching a disclosed $20 billion annualized by the end of 2025 and Anthropic a disclosed $14 billion by February 2026, per Epoch AI's compilation of company disclosures, but they compete at a layer defined by the fastest price deflation ever recorded, under capital requirements David Cahn of Sequoia framed in 2024 as a hole growing from $125 billion toward $500 billion of required end-revenue. The a16z 2025 CIO survey found 37 percent of enterprises already running five or more models, switching deliberately kept cheap. Casado and Wang wrote in August 2025: "To date, there is no model monopoly. And there is unlikely to be in the foreseeable future." A layer with no monopoly, structural deflation, and half-trillion-dollar capital intensity can produce large companies. It is a hard place to produce excess returns.

It will not accrue to raw data hoarding. Sequoia said the quiet part in their own retraction: application data alone "does not create an insurmountable moat." Data earns its corner of the Triangle only when it is proprietary, flowing, and wired into execution and memory, outcome data that improves the work product, not archives that decorate the pitch deck.

And it will not accrue to interface polish. The 14 percent median stickiness statistic is what surface-only value looks like when the novelty tax expires. Layer 7 is where value is delivered. It has never been where value is defended.

A closing note on where the map points next, offered as judgment rather than measurement. Capital is already migrating down the chain toward the physical bottlenecks, Layer minus 1 and Layer 0: power and grid interconnect, advanced packaging and foundry capacity, high-bandwidth memory, the thermal and water constraints that now gate data center construction. Crunchbase data shows roughly 80 percent of first-quarter 2026 venture dollars going to AI, with OpenAI and Anthropic alone absorbing $217 billion, 43 percent of all H1 funding. The market reads this as a bet on models. The chain reads it differently: most of that capital passes through the model companies on its way to compute, which is to say, to the bottleneck layers beneath them. The substrate is being funded at historic scale while the durable application positions, the Triangles and the verification gates, remain comparatively cheap. That mispricing, if the framework is right, is where the next decade's excess returns are hiding.

XI. What to do with this map

For a founder, the framework compresses to a discipline: audit yourself before the market does. Eight questions, scored honestly. How exposed are you to the model layer? What data do you own that cannot be bought? How deep is your workflow control? Do you hold any gatekeeping or trust position? Who controls your distribution? Does your product compound through memory? What does it cost a customer to leave? And if the next platform release shipped your feature tomorrow, what would be left? The classifications run from thin wrapper to useful tool to workflow product to defensible AI system to the rarest position, an intelligence gate. Most honest self-audits land lower than the pitch deck claims. That is the point of running them.

The whole audit can be compressed further, into a single question worth naming: the Absorption Test. Assume the platform beneath you improves by another order of magnitude, on capability and on price, because on the evidence of the last three years it will. Then ask what that does to your position. If the answer is "we become a feature," the position is temporary, whatever the growth curve currently says; Jasper's growth curve was excellent until the week it was not. If the answer is "our data, our playbooks, and our memory become more valuable," because a stronger substrate makes every layer you own work harder, the position is durable. Cursor passes this test: every model improvement makes its orchestration and memory more productive. Chegg failed it. One question, asked before the term sheet, separates the two.

For an investor, the framework replaces a vibe with a checklist. The last three years punished both errors symmetrically: those who funded surfaces as if they were chains lost to Law One, and those who dismissed chains as surfaces missed Cursor, the single largest software return of the era, precisely because the wrapper heuristic had no vocabulary for a company that owned Layers 4, 6, and 8 under a rented Layer 2.

For the industry, the claim of this essay is larger, and I will state it without hedging. The moat debate of the last three years, wrapper versus not-wrapper, model versus application, has been conducted one layer at a time, which is why it keeps ending in contradiction. The unit of analysis was wrong. Competitive advantage in the intelligence economy is not a property of companies. It is a property of positions in a supply chain, exactly as it was for oil, semiconductors, and logistics before this. The firms that internalize this will stop asking whether a company "is AI" and start asking which links it owns, how scarce those links are, and what the four laws will do to every link it rents. Intelligence itself is deflating toward free. The chain that produces, verifies, delivers, and remembers it is where every durable fortune of this era will be built.

The Supply Chain of Intelligence™ and The Intelligence Cube™ are trademarks of Anand Arivukkarasu. Framework: https://supplychainofai.com.

Sources and notes

All quotations are verbatim from the cited texts. Private-company revenue and valuation figures are reported figures from the cited outlets, not audited financials, and are dated in the text. Podcast statements by Martin Casado and YC partners are paraphrased from published episodes rather than quoted, since only machine transcripts were available for verification.

Andreessen Horowitz: "The New Business of AI" (Casado and Bornstein, Feb 2020, a16z.com); "Who Owns the Generative AI Platform?" (Bornstein, Appenzeller, Casado, Jan 2023, a16z.com); "LLMflation" (Appenzeller, Nov 2024, a16z.com); "16 Changes to the Way Enterprises Are Building and Buying Generative AI" (Wang and Xu, Mar 2024); "How 100 Enterprise CIOs Are Building and Buying Gen AI in 2025" (Wang, Xu, Kahl, Erten, Jun 2025); "Questioning Margins is a Boring Cliche" (Wang and Casado, Aug 2025); "What 'Working' Means in the Era of AI Apps" (Moore and Andrusko, Jun 2025); a16z Podcast "Where Value Will Accrue in AI" (Casado and Wang, May 2025) and "Software is Eating Labor" (Rampell, Oct 2025). Andrew Chen, "Revenge of the GPT Wrappers" (Substack, Feb 2025).

Sequoia Capital: "Generative AI: Act Two" (Huang and Grady, Sept 2023); "Generative AI's Act o1" (Huang and Grady, Oct 2024); "AI's $600B Question" (Cahn, Jun 2024).

Y Combinator: PitchBook, "Y Combinator is going all-in on AI agents" (Jun 2025); TechCrunch on the W25 batch's AI-generated codebases (Mar 2025); YC Lightcone Podcast, "Vertical AI Agents Could Be 10X Bigger Than SaaS" (late 2024); YC Requests for Startups (ycombinator.com/rfs).

Bessemer Venture Partners, "The State of AI 2025" (bvp.com). Menlo Ventures, "The State of Generative AI in the Enterprise" (2024 and 2025 editions, menlovc.com).

Market data and cases: Crunchbase News on H1 2026 global funding (Jul 2026); PR Newswire and TechCrunch on Jasper (Oct 2022); Maginative citing The Information on Jasper's markdown (Sept 2023); Fortune, CNBC, and Forbes on Chegg (Oct 2025); ppc.land and The Pragmatic Engineer on Stack Overflow (2025 to 2026); VentureBeat on Devin repricing (Apr 2025); The Next Web and CNBC on Cursor/Anysphere (Nov 2025 to Apr 2026); Forbes, CNBC, Harvey's company blog, and TechStartups citing The Information on Harvey (Oct 2025 to Aug 2026); TechCrunch on Sierra (Nov 2025); Glean press release (Jun 2025); TechCrunch on Abridge (Jun 2025); Crunchbase News on OpenEvidence (Jan 2026); NVIDIA Q1 FY2027 results press release (May 2026); Epoch AI, "Anthropic and OpenAI revenue" data insight (Feb 2026); BenchLM pricing tracker for historical API prices (matching vendor-published price pages); The Futurum Group consolidation survey (Feb 2026); TIME interview with Vinod Khosla (Sept 2024).