The design company everyone calls the AI-native winner is trading 28% below its IPO price while revenue accelerates for the third straight quarter. The incumbent everyone calls the dinosaur has been metering AI usage since 2023 and is quietly re-platforming its sales organization around agents. What each motion actually does, where both companies' filings contradict their investor stories, why one vendor became Figma's partner and competitor and board member and federal supply-chain risk inside a single fiscal year, and the three moves I would hand Figma before its next earnings call

Buried on page 40-something of Figma's most recent 10-Q, in the risk factors nobody reads, is a sentence that describes the company shooting itself.

Read that twice. Figma sells seats. Figma built agent write-access to its own canvas, shipped it in March, and then told the SEC the thing it shipped could reduce how many seats its customers need.

This was not done to Figma by Anthropic or Adobe or anyone else. Figma built it, launched it, and disclosed the consequence in the same fiscal year.

And the escape hatch is welded shut in the same document. Figma says it is "evaluating consumption-based pricing and metering mechanisms" for API and MCP access, then concedes that doing so "may accelerate seat-replacement dynamics rather than offset" them. Charge for the agents and you may just make replacing humans with agents cheaper and more legible. The company has written down that the obvious fix might make the problem worse.

Meanwhile the numbers are, by any normal reading, excellent. Q2 2026 revenue of $370.1 million, up 48% year over year, the third consecutive quarter of accelerating growth. Net dollar retention at 136% and rising. Guidance raised three times in 2026 and never cut. And the stock closed at $23.73 on August 12, down roughly 83% from its second-day peak of $142.92 and about 28% below the $33 IPO price.

The market has repriced the machine underneath the demand while leaving the demand alone.

So I want to pull both companies apart on go-to-market alone, because the received wisdom here is lazy in both directions. Figma is not the obvious AI-native winner, and Adobe is not the obvious dinosaur. And because a teardown that only flatters the challenger is worthless, I will spend real time on where Figma is structurally exposed, including to a company most people think of as its partner.

The arc I keep staring at is the divergence in the middle column. Revenue growth went 38% → 40% → 46% → 48% while the equity fell to roughly a fifth of its peak. Demand held up throughout. What the market marked down was the cost structure and the pricing model sitting underneath that demand, which makes this a go-to-market story rather than a product one.

Six dimensions, each graded on evidence, plus a composite for how AI-native each company's own go-to-market is. Not the product. The motion.

Composite: Figma B+, Adobe B-.

That is the narrowest gap this index has produced in eight teardowns. It is also the first time an incumbent has won a dimension outright. Adobe takes own GTM machinery, and ties Figma on demand creation. I will walk both, because they are the rows that break the pattern.

Figma's motion is genuinely good, and the parts that are good are not the parts people talk about.

Start with what it publishes. Go to figma.com/pricing and you will find a price on every tier including Enterprise: Full seats at $16, $55 and $90 depending on plan, Dev seats at $12, $25 and $35, Collab seats at $3 and $5. I counted the calls to action on the page: fourteen "Get started" against fourteen "Contact sales." An enterprise software company that prints its enterprise price is doing something most of its category refuses to do. When a buyer's first question now goes to a language model rather than a salesperson, a published number stops being a pricing decision and becomes a distribution asset.

Then the community machine, which is larger than its reputation. Figma runs 250+ Friends of Figma chapters across 82 countries that hosted more than 900 events last year, and Config has gone from 1,000 attendees in 2020 to over 10,000 in 2026, with a first-ever Config India landing in Bengaluru in October.

The most interesting thing Figma has built, though, is a file almost nobody reads. I pulled figma.com/robots.txt and it runs to 379,090 bytes and 8,849 lines, carrying 8,763 individual Allow: rules against 37 Disallow: rules. Figma blocks the training crawlers outright, GPTBot, ClaudeBot, CCBot, Google-Extended, cohere-ai, all disallowed at the root. Then it admits the answer crawlers, OAI-SearchBot, PerplexityBot, Claude-SearchBot, ChatGPT-User, against a hand-curated allowlist of thousands of specific URLs, localized into Japanese, German, Spanish and French.

That is not a wall. It is a storefront built for language models. Somebody at Figma decided, URL by URL, what an answer engine is permitted to see, and the answer was: the Config recaps, the Spotify design-system story, the Kimberly-Clark growth story. Block the crawlers that would train a rival, admit the crawlers that answer a buyer, and show them the case studies.

The internal stack matches. A DNS read of figma.com returns Google Workspace with no legacy security gateway, plus Segment, Stripe, Greenhouse, Zendesk, Linear, Wiz, Cursor and Decagon. It carries domain verification for both Anthropic and OpenAI, which corroborates the "model-agnostic architecture" CFO Praveer Melwani described on the Q2 call. And it publishes an MCP public key directly in DNS (v=MCPv1; k=ed25519), which I have not seen another company in this index do.

The pricing hedge is real and it is working. Figma bundled AI credits into every seat, began enforcing limits in March 2026, and by Q2 more than 80% of paid customers above $10,000 ARR were consuming AI credits weekly. Non-GAAP gross margin, which had fallen to 82% in Q1 under inference costs, recovered to 85% on the first full quarter of credit monetization. Melwani's framing on the call was that retention growth now comes from "customers expanding both seats and AI credit add-ons."

Here is the part that complicates the tidy version. Figma's sales organization is completely conventional. The careers page runs SMB, Mid-Market, Enterprise, Strategic and Emerging Enterprise account executives replicated across eight international hubs, quota-carrying, with a live enterprise req targeting 5,000-plus-employee accounts at $165,000 to $190,000 base. That is the same org chart Salesforce had in 2015. Underneath it sits Reachdesk, an account-based direct-mail gifting platform, which showed up in Figma's own DNS records and is about as far from AI-native as a go-to-market tactic gets.

And every AI-in-GTM role Figma is hiring for sits in marketing. Marketing Engineer for AI Deployment. Growth Platform engineering using AI for brand consistency checks and copy generation. Product Manager for AI Growth. There is no agentic function anywhere near the revenue organization.

Adobe's distribution is the most underrated asset in this teardown, and I say that as someone who expected to write the opposite.

Acrobat and Express monthly active users went from 700 million to over 850 million year over year. "Creative freemium" MAU, covering Firefly, Express, Premiere, Photoshop and Lightroom on web and mobile, went from 50 million to over 90 million, up 70%. Management said openly on the Q2 call that they are modeling this on the 1990s Acrobat Reader playbook: give away the reader, monetize the creator.

Then the education pipeline, which compounds on a timescale no venture round can buy. Adobe reports 43 million K-12 students and teachers on free Express for Education, roughly six million higher-education students with campus access, and in January it struck a deal with Airtel to give all 360 million of its Indian customers free Adobe Express Premium for a year. A sixteen-year-old handed Photoshop becomes a twenty-six-year-old who expenses it.

Adobe MAX deserves a specific note because of its economics. It is a paid conference, $1,595 to $1,995 for a full pass, drawing 10,000 to 11,000 attendees. Figma books Config as a sales and marketing cost that compressed its operating margin last quarter. Adobe books MAX as revenue. Same motion, opposite sign.

On monetization Adobe is not behind, it is ahead. It has run generative credits since September 2023, more than two years before Figma enforced a single credit limit, and in February 2026 it relaunched Firefly with unlimited standard generations while metering only premium video and audio models. That is a more refined free-versus-paid split than Figma's flat monthly allocation. Separating the marginal cost of generation from the fixed cost of tooling is the economically correct architecture, and Adobe built it inside a forty-year-old per-seat licensing company without breaking the base subscription. Most incumbents cannot do this at all.

And then the finding I did not expect. Adobe is investing more in agentic go-to-market than Figma is. Its live job postings include a "GTM AI & Agentic Products" team building "an always-on Agentic operating layer that connects strategy, data, and frontline execution," a separate "GTM Agentic AI Operations" team, a Director of AI GTM Strategy & Innovation at $177,000 to $323,000, and a Principal AI Strategist whose posting says the job is to "audit and unify grassroots AI initiatives already underway" in the sales organization and map them to pipeline, win rates and net revenue retention. That last phrase describes adoption that has already happened, not a hiring wish.

Now the break, and it is severe.

In March 2026 Adobe settled with the FTC for $150 million, $75 million in civil penalties and $75 million in customer credits, over how it sold and cancelled subscriptions. The original complaint named two executives personally. The mechanic at issue is still live on Adobe's own help page today, in Adobe's own words: "After 14 days, a cancellation fee (early termination fee) of 50% of the remaining balance of the contract applies."

The customer sentiment is exactly what you would predict. Adobe's Trustpilot rating is 1.2 out of 5 across 7,374 reviews, against 4.5 on G2. Trustpilot's own summary of that corpus names the mechanism without prompting: "People frequently struggle with unexpected subscription charges and difficult cancellation processes that involve very high exit fees. Consumers also mention being automatically enrolled in extra plans or free trials without clear disclosure." That is the FTC complaint, restated by customers, two years later. The G2 gap is a segmentation rather than a contradiction: G2 is where professionals rate the software, Trustpilot is where people go after being charged $300 to leave. An entire genre of "how to cancel Adobe without the fee" guides exists, documenting a plan-change workaround. When your customers are engineering exploits against your retention mechanics, that is a go-to-market finding, not a support ticket.

The second break is stranger, and it is the one I cannot stop thinking about. Adobe paid $1.9 billion for Semrush, closing in April, and justified it explicitly as a bet on generative engine optimization as the successor to SEO. Then I pulled adobe.com/robots.txt. It is 10,199 bytes and contains zero AI crawler directives. Not one. A grep for GPTBot, ClaudeBot, Google-Extended, PerplexityBot, Anthropic, OpenAI or Cohere returns nothing at all. What it does declare is gsa-crawler-www, the crawler for the Google Search Appliance, a product Google discontinued in 2019, and Atomz/1.0, a site-search tool acquired in 2005.

Adobe spent $1.9 billion to help other companies get found by AI, and has not configured its own front door for it.

Except, and this is the honest complication, Adobe is winning that game anyway. Because it never opted out, three decades of Adobe tutorials, help documentation and forum threads went into the training corpus of every frontier model unblocked. Adobe is the default answer when you ask a model how to remove a background. Figma, which blocked ClaudeBot and GPTBot from training, is systematically less present in the weights that now mediate an enormous share of software discovery. Figma engineered an 8,763-rule allowlist so answer engines could see some marketing pages. Adobe did nothing and got its whole corpus into the models.

Both companies are also missing the same number, twice. Adobe reports 43 million students and 90 million freemium users and has never published a conversion rate for either. Reach without conversion, disclosed twice, is a tell.

Adobe's flagship AI metric is roughly 2% of its business, and the pivot chasing it cost about what the metric is worth. Adobe disclosed AI-first ARR "tripled year over year to more than $500 million." Against total ARR of $27.10 billion, that is about 1.8%. To grow that slice, management deferred planned Creative Cloud price increases and pivoted to freemium, at a cost analysts put near $500 million in near-term organic ARR. The company gave up roughly as much organic ARR as the entire AI book it was showcasing. Only the growth rate reaches the slide.

Worth noting Adobe runs two AI metrics it defines itself: "AI-first ARR" at ~$500 million and "AI-influenced ARR" at over $5 billion. A self-defined metric with a tenfold gap between its narrow and broad readings is an instrument, not a measurement.

"No material AI disruption" sits directly beside a pricing retreat. HSBC upgraded Adobe in July on the view that "there is little evidence that emerging AI-focused rivals are taking meaningful market share." On the narrow claim, HSBC is right, and I want to be fair about that: nothing in this corpus shows an actual seat-count or customer-count decline caused by an AI competitor. Revenue growth has ticked up, 11% to 12% to 13%. But a company feeling no pressure does not shelve a planned price increase and redirect hundreds of millions into a free tier, and Adobe's own President said the shift "is playing out more quickly than we had planned for." The trailing data says fine. The capital allocation says worried.

One analyst reversed himself twice on the same name. Morgan Stanley's Keith Weiss upgraded Adobe to Overweight in 2023 at a $660 target, citing "clarity on AI-enabled products and the monetization roadmap." By September 2025 he cut to Equal-Weight, saying there is "relative uncertainty in a sizable portion of the Adobe base where we lack confidence in Gen AI advancements being a net positive." He has since gone to Underweight with a $240 target. That is harder to dismiss than two banks disagreeing.

And on Figma's side, the growth engine and the risk factor are in the same day's filings. Melwani's press release says retention is strong "as customers expanded both seats and AI credit add-ons." The 10-Q filed the same day says third-party agents "could reduce the number of paid seats required for a given level of output." Both documents are true. They just cannot both be the future.

Adobe's own 10-K says the same thing about model vendors, and says it has already happened. Buried in the competition risk factor: Adobe expects "more competition as AI continues to advance," and specifically "from companies offering generative and agentic AI solutions, including but not limited to prompt-based and multi-modal creation and editing." That is a description of Claude Design and Figma Make without naming either. Then the line that matters: "Other companies have in the past, and may in the future prevent, limit or interfere with our ability to use third-party models in our solutions."

Have in the past. Both companies in this teardown have written into their filings that their AI roadmap runs on models they do not control, and both have said the arrangement has already caused them a problem. Figma names its vendor. Adobe does not.

Here is the section that matters, and it is not about Adobe at all.

Inside one fiscal year, Anthropic became four things to Figma simultaneously.

A partner. On February 17, Figma shipped "Code to Canvas," a flagship integration with Claude Code that captures a running UI into editable Figma layers through Figma's MCP server.

A competitor. On April 17, nine weeks later, Anthropic shipped Claude Design, which turns prompts into prototypes, reads your codebase to apply your design system, and exports to Canva. It is bundled into Claude Pro at $20 a month with no separate charge. Figma's Full seat runs $16 to $90. A model vendor can give away the adjacent application because the subscription is the business. FIG fell 6 to 7% that day; Wix fell 4.7%, GoDaddy 3%, Adobe 2.7%.

A governance problem. Anthropic's Chief Product Officer, Mike Krieger, sat on Figma's board and resigned on April 14. Activist investor Findell Capital wrote to the board on May 28 calling the sequence a "serious corporate governance concern" and asking for an independent investigation into "whether Anthropic benefitted from any improper use of Figma's confidential information." I want to be precise here, because the activist's framing is tidier than the record: Krieger resigned the same day The Information reported that Anthropic's next model would ship design tools, and TechCrunch's headline was "Anthropic CPO leaves Figma's board after reports he will offer a competing product." That reads as a director stepping down when a conflict became public. Findell is also arguing Figma is undervalued and pushing for cost cuts, so it has an interest in board pressure.

And a federal supply-chain risk. This one is in Figma's own 10-Q, and it is the strangest sentence in the corpus:

The directive is enjoined and under litigation, and Figma's filing is careful to say so. But the exposure is real: Figma built its government AI features on one vendor's model, and that vendor is currently the subject of a federal directive.

Four exposures. One counterparty. And Figma's robots.txt blocks ClaudeBot from training on its corpus while its government product runs on Claude.

And before this reads as a one-sided hit, Figma has a version of Adobe's problem too. Its own Trustpilot sits at 2.1 out of 5 across 228 reviews, and Trustpilot's summary of that corpus could have been written about Adobe: "Many individuals experienced unexpected charges, feeling that the subscription pricing model is confusing and overly expensive for standard users. Consumers also pointed out persistent difficulties with trying to downgrade plans." Figma's 10-Q concedes the same thing in filing language, reporting that after it began enforcing AI credit limits in March 2026 it "observed elevated customer support volume, instances of customer dissatisfaction expressed through public and social channels, and reduced usage by certain customers." Smaller scale, no regulator, same failure mode: a billing system the customer cannot predict.

The smaller fragility, which will bite sooner: Figma's sales and marketing line now contains its own compute bill. From the 10-Q, Figma classifies "within sales and marketing technical infrastructure and hosting costs, including AI inference... related to the users of our free version of Figma." Every analyst reading a 58% S&M increase is reading it as sales headcount. It is partly the cost of serving free users AI, plus $19.0 million of post-IPO stock compensation out of a $57.2 million increase. The free tier is now a metered expense that scales with usage, and it sits in the line item that is supposed to prove efficiency.

Trigger: the 10-Q says Claude is "the large language model on which we have built the AI features embedded in our governmental offerings," while a federal directive against Anthropic sits enjoined and in litigation, and Melwani has separately described a "model-agnostic architecture."

Why it moves the needle: the architecture claim and the disclosure claim currently contradict each other in public, and the buyer who cares most, a federal procurement officer, reads the 10-Q. Figma already has OpenAI domain verification live in its own DNS, so the second path plausibly exists. Naming it converts a risk factor into a procurement answer. This is a one-page disclosure and a press note, not an engineering project.

Trigger: figma.com/robots.txt runs 8,763 Allow: rules to admit answer engines while blocking GPTBot, ClaudeBot, CCBot and Google-Extended from training, and Adobe, which blocks nothing, is the default model answer in this category by inheritance.

Why it moves the needle: the allowlist optimizes retrieval, which affects the answer a model gives when it already knows about you. Training presence affects whether it thinks of you at all. Figma is currently paying engineering cost for the second-order lever while ceding the first-order one to a competitor that made no decision. The narrow version is defensible: keep the block on product surfaces and community files, lift it on the customer stories and Config content that are already in the retrieval allowlist. Those pages have been judged safe to show a model. There is no coherent reason they are unsafe to train one.

Trigger: the 10-Q discloses that free-user AI inference is booked inside sales and marketing, and S&M grew 58% against 48% revenue growth, of which $19.0 million was post-IPO stock compensation.

Why it moves the needle: the single loudest bear argument on Figma is that go-to-market efficiency has inverted. That reading is wrong. Figma holds the numbers to disprove it, buried in a line item that invites the opposite conclusion. Splitting free-tier inference out reframes a sales-inefficiency story as a customer-acquisition-cost story with a falling unit cost curve. The stock is down 83%; the cheapest available repricing catalyst is an accounting disclosure the company already computes internally.

Three things travel out of this.

Your model vendor's roadmap is your competitive roadmap, and you will learn it on their schedule. Figma shipped its Claude Code integration in February and met Claude Design in April. If your AI features run on somebody else's model, you have a partner who can enter your category at zero marginal cost, priced inside a subscription they already sell.

A pricing model is a bet on what stays scarce. Per-seat pricing bets that humans holding licenses remain the unit of value. Both companies here know it and both have hedged with usage metering; the incumbent started two years earlier, which is not what anyone expects when they open a teardown like this.

And the grade you deserve is not the grade the narrative gives you. Adobe runs the better-instrumented monetization architecture, the larger compounding distribution asset, and more agentic go-to-market investment than the company the market calls AI-native. It also settled with a regulator for $150 million over how it keeps customers from leaving. Both facts are true, they land in different rows, and any story that averages them into "legacy incumbent" is not analysis.

The company that shipped the feature that eats its own seats is the one being punished. The company that never updated its robots.txt is the one the models recommend.