Every version of the AI capex debate starts with training runs. The other half is inference billed to professionals, one task at a time, and this week both halves met: Adobe put its Creative Agent into five flagship apps on Nov. 10 and previewed third-party agents operating on native Photoshop and Premiere files, while Anthropic — one day earlier — launched flagship models at token prices 20% below the previous generation. Adobe's cost of revenue shows the flow arriving: subscription cost of revenue rose 24% year over year in fiscal Q3 2026, with hosting services and data-center costs up 22%.
Adobe AI-first ending ARR
>$650M
Q3 FY2026, quarter ended Aug. 28, 2026, growing more than 150% year over year
Firefly ARR
+40% QoQ
Q3 FY2026, across the Firefly app and credit packs, with credit consumption accelerating sequentially
Claude Fable 5.2 output price
$40 / 1M tokens
Launched Nov. 9, 2026, 20% below Fable 5.1; fast mode priced at $80
Adobe subscription cost of revenue
+24% YoY
Q3 FY2026, with hosting services and data-center costs up 22%
Anthropic compute commitments
$895.4B
Amended IPO prospectus, Nov. 2, 2026, against nine-month revenue of $33.2B
Adobe ending ARR
$27.5B
Q3 FY2026, up 11.2% year over year, with remaining performance obligations of $22.16B
The launch
Adobe Turned Five Flagship Apps Into Agent Surfaces in a Single Morning
Adobe expanded the Adobe Creative Agent across Firefly and Creative Cloud on Nov. 10: AI Assistants powered by the agent are in beta in Premiere, Illustrator, InDesign, After Effects and Frame.io, and photographers get the same assistant inside the Photoshop beta app. The more consequential piece is a preview of local Model Context Protocol integrations for Photoshop, Illustrator, Premiere and After Effects, where third-party agents make changes directly in native project files on the user's own machine while layers, vectors and sequences stay editable. Adobe says third-party agent support spans frontier models including OpenAI's and Anthropic's, and that it is working to bring Adobe's creative technology to ChatGPT, Claude, Gemini and Copilot as plugins.
| Capability | Where it runs | Availability |
|---|---|---|
| Creative Agent AI Assistants | Premiere, Illustrator, InDesign, After Effects, Frame.io | Beta, Nov. 2026 |
| Creative Agent AI Assistant | Photoshop beta desktop app | Beta, Nov. 2026 |
| Local MCP integrations for third-party agents | Photoshop, Illustrator, Premiere, After Effects | Private beta, early 2027 |
| Always-on Firefly agents (asset triage, quality checks) | Firefly | Preview |
| Firefly Design, Firefly Graph, Firefly Builder | Firefly | Beta, beta, private beta |
| Premiere Generative Media (video, sound effects, soundscape, music) | Premiere | Models from Google, Kling, Runway, Luma plus Firefly |
| Relight, Generative Restyle, AI Assisted Editor | Photoshop | Generally available and beta |
| CX Enterprise Coworker inside GenStudio | GenStudio and Experience Cloud | Generally available since June 2026, 1,700+ early adopters |
- One edit now multiplies: Adobe's own example removes a logo once and cascades the change across a campaign, turning a single instruction into dozens of billable generations.
- Premiere's new agent capabilities apply an effect across every clip in a sequence, track a masked subject through a frame and grade a whole sequence at once, multiplying inference volume for a single edit.
- Third-party agents operate on the user's machine against native project files, so Adobe keeps the seat and the file while the generation bill goes to whichever model provider supplies the agent.
- Every generation is priced through Adobe's own generative-credit meter, so the credit card, not the subscription, is where agentic usage becomes revenue.
Pricing
A 20% Price Cut Only Pays If Volume Grows Faster
Claude output token prices, launch price per million tokens
Anthropic's published launch prices: Opus 5.5 on Sept. 22, 2026, Haiku 5.5 on Oct. 7, Fable 5.2 on Nov. 9.
Unit: $ per million output tokens
Haiku 5.5
Small-model output, used for routine agent steps
0.5
Opus 5.5
Down from Opus 5's $25, a 20% cut
20
Fable 5.2
Down from Fable 5.1's $50, a 20% cut on Nov. 9
40
Fable 5.2 fast mode
Optional generation at up to 2.5x speed
80
Anthropic cut flagship output prices twice in a quarter — Opus 5.5 to $20 per million tokens from $25 in September, then Fable 5.2 to $40 from $50 on Nov. 9 — and halved Sonnet 5.5 cache reads to $0.10 per million, which the company says makes Sonnet 5.5 roughly 20% cheaper on most agentic work. A 20% price cut destroys revenue if volume is flat: the dollar line only holds if the same workload grows by about a quarter. Agentic creative work is the volume source that keeps that arithmetic intact, because one instruction fans out into many generations, each carrying the context of the file being edited.
Unit economics
Adobe's Meter Is Running, and It Is the Smallest Line in the Stack
Two numbers frame Adobe's agentic experiment. AI-first ending ARR above $650 million is growing more than 150% year over year, but that is only about 2.4% of $27.5 billion of total ending ARR, so the agent pivot has not yet changed what Adobe is sold on. Total ARR growth is 11.2%, and the fiscal 2026 revenue target of $26.576 billion to $26.626 billion implies a business still growing at low-teens rates.
| Metric | Q3 FY2026 | Read-through |
|---|---|---|
| Revenue | $6.76B | Total ending ARR $27.5B, up 11.2% year over year |
| AI-first ending ARR | >$650M | Up more than 150%, about 2.4% of total ARR |
| Firefly ending ARR | +40% QoQ | Across the Firefly app and credit packs; credit consumption accelerating sequentially |
| Creative freemium monthly actives | >100M | Growing more than 70% year over year |
| Subscription revenue | $6.58B | Up 14%, against subscription cost of revenue up 24% |
| Gross margin | 88.7% | Cost of revenue $763M, versus 89.3% for fiscal 2025 |
| Remaining performance obligations | $22.16B | Up 8% year over year |
| FY2026 revenue target | $26.576B–$26.626B | Ending ARR book of business growth of 10.2% |
The cost side is where the agents already show up. Adobe's filing attributes the increase in subscription cost of revenue to hosting services and data-center costs, which rose 22%, and states that subscription costs include AI inferencing costs. Subscription revenue grew 14% over the same period, so inference is now a cost line in Adobe's income statement, arriving ahead of the revenue it is supposed to create.
Competitive pricing
Adobe's Credit Card Is a Published Price List for Video Inference
| Partner model | Credits per second, 1080p | Also available in | Note |
|---|---|---|---|
| Kling 2.5 Turbo (Kuaishou) | 15 | Firefly, Firefly Boards | Cheapest per-second rate in the list |
| Kling 3.0 (Kuaishou) | 25 | Firefly, Firefly video editor, Boards | Text to video and image to video |
| Runway Gen-4 (Runway) | 40 | Firefly Boards | 720p priced at 20 credits per second |
| Veo 3.1 (Alphabet) | 50 | Firefly, Premiere Generative Media | Veo 3.1 Fast priced at 20 |
| Luma Ray3 (Luma AI) | 125 | Firefly, Premiere Generative Media | 4K priced at 150 credits per second |
Those credit rates are a price list for inference that Adobe publishes to its own users, and the spread is the competitive battlefield: the same second of 1080p video costs 25 credits on Kling 3.0 and 125 on Luma Ray3, a fivefold difference. Whoever wins the agent loop wins the meter, because the agent decides which model generates each frame, clip and campaign variant, and Adobe's credit card decides what that decision is worth. What Adobe does not disclose is the cost of serving a credit, so the take rate on agentic work cannot be computed from public filings.
Supply chain
Who Actually Gets Paid When an Agent Edits a Video
| Layer | Named player | Evidence |
|---|---|---|
| Models and agents | Anthropic and OpenAI | Adobe's MCP preview names frontier models including OpenAI's and Anthropic's; Adobe is also bringing its creative technology to ChatGPT, Claude, Gemini and Copilot as plugins |
| Model labs' compute | Anthropic | About $895.4 billion of compute and infrastructure purchase commitments against nine-month revenue of $33.2 billion |
| Model capital | NVIDIA, Alphabet, Microsoft and four other investors | $26 billion private placement in Anthropic stock at the IPO price; offering priced Nov. 18, Nasdaq debut Nov. 19 |
| Accelerated computing | NVIDIA | Adobe will build its next Firefly foundation models on NVIDIA libraries, with Agent Toolkit and Nemotron explored for hybrid, long-running agent loops |
| GPU cloud | CoreWeave | Second-quarter revenue $2.575B, revenue backlog near $104B, more than $25B of net-new commitments in early Q3, 1.5GW of active power |
| Media inference | fal.ai | fal models made available to Adobe Express and Firefly users |
| Hyperscale clouds | Amazon and Alphabet | AWS trains Adobe's Firefly models and Bedrock supplies partner models; Google's Veo 3.1 is priced inside Adobe's own meter |
| Application layer | Adobe | Creative Agent in five flagship apps, more than 100M freemium monthly actives, and CX Coworker used by 1,700+ customers |
Who is actually writing the checks for AI inference
Capital expenditure over the trailing twelve months, from each company's latest quarterly filing.
Unit: $ billions of capital expenditure, trailing twelve months
Amazon
Up from $131.8B in fiscal 2025
215.2
Alphabet
Up from $91.4B in fiscal 2025
159.8
CoreWeave
Trailing free cash flow of negative $13.7B
20.6
NVIDIA
Seller of the hardware, not the biggest buyer
7.4
Adobe
$214M, less than 1% of revenue
0.2
The cash behind that chain sits with the buyers, not the seller. On a trailing-twelve-month basis, Amazon spent $215.2 billion on property and equipment, Alphabet $159.8 billion, CoreWeave $20.6 billion and NVIDIA $7.4 billion, while Adobe spent $214 million. Adobe is the pass-through in the AI capex cycle, metering usage it neither owns the silicon nor owns the models behind, which is exactly why its cost line moves before its revenue line does.
The financing structure makes the loop explicit. Anthropic's amended prospectus put committed compute and infrastructure spend at about $895.4 billion against nine-month revenue of $33.2 billion, and seven investors, including NVIDIA, Alphabet and Microsoft, agreed to buy $26 billion of its stock at the IPO price, with the offering priced Nov. 18 at $70 to $78 a share for a fully diluted value of up to $1.405 trillion. That makes the capex debate a financing question as much as a demand question, because the buyers of the compute are now shareholders in the labs that consume it.
Risk
The Unresolved Question: Do Outside Agents Fill Adobe's Meter or Drain It
The MCP preview cuts both ways. Adobe's local integrations run on the user's own machine against native project files, which means an agent from another model provider can act on Adobe's canvas while the generation is billed to that provider's account rather than to Adobe's generative-credit meter. Adobe has not disclosed how third-party agent generations will be charged, and if the credit meter is bypassed, the same feature that creates demand disintermediates the meter Adobe is building to monetize it.
- Undisclosed: whether generations executed by third-party agents bill to Adobe credits or to the agent's own model account, which leaves the take rate on agentic work unverifiable.
- Undisclosed: Adobe reports Firefly ARR growth of 40% sequentially but not token volumes, tasks per agent or the cost of serving a credit, so no investor can compute inference elasticity from the filing.
- Undisclosed: Anthropic's prospectus reports revenue and commitments but not token volumes, leaving the price-versus-volume trade off invisible.
Horizons
What to Watch in Days to Quarters, and What Decides the Next Three Years
- Early 2027: local MCP integrations enter private beta for Photoshop, Illustrator, Premiere and After Effects, the first test of whether outside agents create usage or displace it.
- December 2026: Adobe's fiscal fourth-quarter results against a $26.576 billion to $26.626 billion revenue target, with credit consumption and hosting costs the two lines that matter.
- Nov. 18 to Nov. 19: Anthropic prices and lists on Nasdaq with $895.4 billion of compute commitments behind it, putting a public market price on agentic inference demand.
- Near term: CoreWeave's $104 billion backlog and more than $25 billion of early-third-quarter commitments convert into served revenue, showing whether booked inference demand is real capacity.
- One to three years: AI-first ARR has to move from about 2.4% of $27.5 billion toward a visible share of the base, or Adobe remains a seat business with a cost problem.
- One to three years: the partner-model mix inside the credit meter decides who owns the inference margin, with Kling, Veo, Runway and Luma all bidding for the same agent loop.
- One to three years: token prices keep deflating while agent loops keep expanding, so compute demand must grow faster than 20% a year just to hold the dollar line.
- One to three years: the $375.0 billion Amazon and Alphabet spent together on capital expenditure has to be earned back by agentic workloads, which makes professional creative agents a marginal buyer rather than a rounding error.
Synthesis
Facts, Inference and the One Number Nobody Has Disclosed
Facts: Adobe shipped agent surfaces into five flagship apps on Nov. 10, third-party agents will operate on native project files in early 2027, Anthropic cut flagship token prices 20% on Nov. 9, and Adobe's subscription cost of revenue rose 24% year over year. Inference: the agentic creative workflow is the marginal volume that makes falling token prices compatible with rising compute spending, and Adobe's credit meter is where that volume is priced. Speculation: that Adobe keeps the take rate once outside agents run inside its canvas, which the filings do not answer. No investor can yet verify the take rate on agentic work, and that is the number the next two disclosure cycles will produce.
Where the agentic creative dollar lands
- AI-first ARR above $650M is only about 2.4% of $27.5B ARR, so agents have not yet changed what Adobe sells.
- Subscription cost of revenue rose 24% while subscription revenue grew 14%, so inference inflation reaches gross margin first.
- Firefly ARR grew 40% sequentially with credit consumption accelerating, which is the clearest sign the meter is filling.
- Over one to three years, local MCP agents running on users' machines could route generation elsewhere and leak Adobe's per-credit take rate.
- Adobe will build its next Firefly foundation models on NVIDIA's stack, tying Adobe agent hours to GPU demand.
- Trailing revenue of $303.0B at a 74.7% gross margin funds $7.4B of annual capital spending while workloads expand.
- NVIDIA's $26B Anthropic placement at the IPO price converts model demand into equity demand.
- Over one to three years, agentic editing loops rather than chat completions become the marginal source of inference hours.
- Revenue backlog near $104B plus more than $25B of early-third-quarter commitments means inference demand is booked years ahead.
- Active power of 1.5GW against 3.7GW contracted gives the capacity to absorb media-generation bursts.
- Balance-sheet risk: trailing free cash flow of negative $13.7B and net debt at 11.5x EBITDA make utilization the whole thesis.
- Gemini connectors in Acrobat and DeepMind work on Photoshop editing give Google a distribution channel into professional creative seats.
- Veo 3.1 at 50 credits per second of 1080p video is priced inside Adobe's meter, so model revenue rides Adobe's agent loops.
- Trailing capital spending of $159.8B, up 75% from $91.4B in fiscal 2025, is the cash side of the same demand Adobe creates.
- AWS trains Adobe's Firefly models and Bedrock serves partner models, so agentic Adobe traffic can be billed on AWS compute.
- Trailing capital spending of $215.2B, up 63% from $131.8B in fiscal 2025, is the largest single bet that volume offsets falling token prices.
- Risk: Adobe's Firefly models are moving to NVIDIA's stack and Google DeepMind, shifting Adobe's own inference spend away from AWS.
- Kling 3.0 at 25 credits per second of 1080p video undercuts Runway's 40 and Luma's 125 inside Adobe's own price list.
- Premiere's Generative Media tool now ships Kling alongside Google, Runway and Luma, giving Kling enterprise distribution it could not buy.
- Risk: Adobe's own Firefly Video competes in the same meter, and partner-model availability varies by region.
