Did FOTOhub have model orchestration before Adobe?

Yes. FOTOhub.app launched on September 11, 2025 with the multi-model, one-subscription, one-credit-balance architecture as its founding premise, twelve months before Adobe embedded five competing engines in the Premiere timeline on September 8, 2026. Founder Mateusz Ulewicz published the underlying thesis - that AI tool fragmentation, not model quality, is the real cost problem - on March 15, 2026, roughly six months ahead of Adobe's release, and shipped the named routing layers FOTOcore AI and Gabriel AI during 2026.

How is FOTOhub's orchestration different from Adobe's model picker?

Three differences of substance. Scope: Adobe's picker covers video inside a professional editing timeline; FOTOhub routes across image, video, audio, chat, text, 3D and virtual try-on, plus storage, automation, an API of 314 endpoints and a 28-tool MCP server, on one credit balance. Automation: Adobe's user chooses from a list; Gabriel AI classifies intent and selects the model, with automatic fallback when a provider errors or refuses a request. Vertical integration: FOTOhub owns models underneath its own orchestrator - IDA Q 1.0 for image, IDA Voice for speech, IDA Y for audio - so the routing layer has a cost anchor no pure aggregator has.

Why does a venture fund care about orchestration rather than models?

Because a model leaderboard position can change in weeks, while a layer embedded in a customer's daily workflow changes slowly. Model count is not a moat and neither is API access; anyone with an integration budget can add another endpoint. What compounds is routing quality on real traffic, workflow retention, proprietary user context, cost control, compliance and distribution. DGE3 underwrote FOTOhub in February 2026 on that basis and has committed $10.8 million; Adobe's September 2026 release moved the market's question from whether an orchestration layer is needed to who will build a good one.

What has not been proven about the orchestration thesis?

Once orchestration goes mainstream it stops being a differentiator on its own, and the model catalogue is the easiest part of the product to copy. Adobe owns the professional editing timeline and a company like FOTOhub will not take that position by frontal attack. Routing quality has to be proven on real production traffic rather than asserted, cost has to stay predictable, and the interface has to stay comprehensible as functionality grows. FOTOhub also does not yet hold SOC 2 or ISO 27001 certification, which is the next gate for enterprise and public-sector contracts.

Adobe just validated the thesis we underwrote: why DGE3 bet on orchestration, not models

Q: What is AI model orchestration?

AI model orchestration is the software layer above the models that decides which model should perform a given task and then executes it in context. Real orchestration has six components: intent classification (what does the user actually want), a capability registry (which models support this resolution, duration, format, reference input or audio mode), routing (choose by quality, cost, latency, availability and data policy, with a fallback if the provider fails), shared context (references, brand, style and prior decisions travel with the job), observability (which model and version ran, what it cost, how long it took, where the output is stored) and workflow portability (a model can be swapped without rebuilding the product). A page with a dozen buttons wired to a dozen APIs is a model catalogue, not orchestration.

Q: What did Adobe announce on September 8, 2026?

Adobe brought generative video and sound effects directly into the Adobe Premiere timeline, with a choice of engine: its own Firefly plus partner models from Google Veo, Kling, Runway and Luma AI. A user marks a gap in the sequence, describes the missing shot, can supply reference frames from the current edit, picks a model and receives an editable clip in place. Premiere shows the credit cost of the selected model and settings before generation runs. Adobe is trading exclusivity over the engine for control of the workflow.

Q: What are Gabriel AI and FOTOcore AI?

FOTOcore AI is FOTOhub's internal layer that manages workflows, resources and routing across the model catalogue. Gabriel AI is the interface above it: it classifies user intent, recommends or selects a model, sets the parameters and can launch multi-step workflows across image, video, music, editing, chat and 3D. Gabriel also runs on FOTOhub's own GPU fleet, which is why it can serve as the free tier's chat model at no marginal token cost.

Q: Is there a best AI model?

Not in practice. Any real project has more dimensions than quality alone: cost, latency, clip length, aspect ratio, audio support, API availability, character consistency, editability, data policy, licensing, processing location and vendor predictability. An analysis of 80 models published in early September 2026 found one image-quality leader in its dataset but eight separate offerings on the price-to-quality frontier - the highest quality score did not produce the best economic choice. The useful question is not which model is most powerful but which model is good enough for this task, at this budget, deadline, format, risk profile and set of legal requirements.

2026-09-13 - Dmitri Evanov

On September 8, 2026, Adobe shipped a feature that reads, to most of the market, like one more generative button. In the latest Premiere you mark a gap on the timeline, describe the shot you are missing, optionally feed it reference frames from the current edit, and choose an engine: Adobe's own Firefly, or Google Veo, Kling, Runway or Luma. The clip lands in the sequence, editable, in place.

Three days later, on September 11, 2026, FOTOhub.app turned one year old — a company built from its first day on precisely that architecture: many interchangeable models, one environment, one context, one balance, a layer above the engines that decides which engine to use.

We are FOTOhub's lead institutional investor. We committed $10.8 million across three tranches beginning in February 2026, and the single sentence at the top of our investment memo was that the durable asset in generative AI is not the model, it is the layer that chooses the model. When the most important company in professional creative software concedes the same point in public, a fund should say what that does and does not mean. This post is that accounting — including the parts that are uncomfortable for us.

The call, and its timestamps

The claim we want to make precisely, because imprecise versions of it are worthless: this architecture was FOTOhub's founding premise, not a reaction to Adobe. The public record is checkable.

Date	What happened	What the market believed at the time
Sep 11, 2025	FOTOhub launches in Bydgoszcz, Poland — multiple models, one subscription, one credit balance, one asset library	The race is between generators. Pick the best model and build your workflow around it.
Feb 2026	DGE3 leads the $6.1M Seed, underwriting the orchestration layer rather than the model catalogue; 460,000 users	Model aggregators are thin wrappers over other people's APIs.
Mar 15, 2026	Mateusz Ulewicz publishes AI Tool Fragmentation: Why You Pay for 10 Subscriptions with Nothing to Show for It — the argument that the binding cost is coordination, not capability, and that platforms should be judged on workflow completeness rather than isolated features	Fragmentation is a user inconvenience, not a market structure.
2026, through the year	FOTOcore AI (workflow, resource and routing management) and Gabriel AI (intent classification, model selection, multi-step workflows) ship into production; catalogue reaches 226 models	Routing is a text-LLM concept; creative work is chosen by hand.
Jul 2026	IDA Q 1.0 — FOTOhub's own image model, trained and served on its own GPU fleet — goes live underneath its own orchestrator	You are either a model lab or an application. Not both.
Sep 8, 2026	Adobe embeds Firefly, Google Veo, Kling, Runway and Luma in the Premiere timeline, with the credit cost of the chosen model shown before the job runs	—
Sep 11, 2026	FOTOhub's first birthday: 1,000,000+ users in 46 countries, 226 models, #265 in the world on Crunchbase CB Rank; its founder ranked #1 globally on CB Rank (Person)	The question is no longer whether the orchestration layer matters.

Twelve months separate the architecture from its validation by the incumbent; the written public thesis precedes Adobe's release by roughly six. We are not going to claim clairvoyance on a founder's behalf — Adobe's implementation is not a copy of anything, and the two companies are not comparable in scale, resources or product history. What we will claim is the thing that matters to an investor: the architectural direction was chosen deliberately, early, and against the consensus of the time, and the consensus has now moved.

What Adobe actually gave up, and what it bought

Read the Premiere release as a trade rather than a feature and its logic becomes obvious. Adobe surrendered engine exclusivity. It could have locked its users into Firefly; instead it put four competitors on the same menu.

Engine on the Premiere menu	Its role in Adobe's ecosystem
Adobe Firefly	Adobe's own family, positioned around production use, data provenance and commercial safety
Google Veo	Partner video engine — additional technical choice without leaving the edit
Kling	Partner model positioned around multi-shot generation, motion and synchronised audio
Runway	Specialist external engine, reachable without export and re-import
Luma AI	Another aesthetic profile and capability set inside the same workflow

In exchange for exclusivity, Adobe took the decision point — the place where a professional supplies context, generates, compares, revises and assembles. It no longer has to win every benchmark. It only has to be the room where the work happens.

That is a stronger position than a leaderboard placing, and any investor who has watched a model ranking turn over in six weeks understands why. The top model changes monthly. An environment embedded in someone's working day changes on a scale of years. Adobe also, tellingly, shows the credit cost of the selected engine and settings before generation runs — the same design decision FOTOhub made across its entire surface, and a decision that only a platform which owns the billing layer can make.

Orchestration is not a model catalogue

Here is where we spend most of our diligence time, because this is where the category will separate. Wiring a dozen buttons to a dozen APIs takes a competent engineer a fortnight. Real orchestration has six layers, and each one is where a thin aggregator fails.

Layer	What it has to do	FOTOhub's implementation as at September 2026
1. Intent classification	Work out whether the user wants to generate, edit, animate, narrate, score, or do several of those in sequence	Gabriel AI classifies intent, recommends the function and can launch multi-step workflows across image, video, music, editing, chat and 3D
2. Capability registry	Know which model supports which resolution, format, duration, reference input, editing method, audio mode or control type	226 models catalogued across eight categories with per-model capability metadata; discovery exposed publicly through the API
3. Routing	Select on quality, cost, latency, availability, data policy and user preference — and fall back when the primary provider fails	FOTOcore AI manages routing and resources; degraded models are pulled from rotation by continuous health monitoring rather than accepting work they cannot finish
4. Shared context	Carry references, brand, project style, prior decisions and technical parameters into every job instead of starting cold	One asset library across all thirteen studios; reference trays, masks and keyframes travel between modalities; the output of any model is the input of any other
5. Observability	Record which model ran, in which version, at what cost, how long it took, what data was sent and where the result is stored	Price quoted before execution on every operation and every surface; per-operation billing records; customer-selectable storage region across 16 AWS regions; public status page
6. Workflow portability	Survive a model being withdrawn, repriced or overtaken without rebuilding the product	One job contract shared by app, REST API and MCP server; adding or swapping an engine is a configuration and adapter change, not an architectural one

IDC, which we cite because it is the analyst framing our LPs already read, recommends designing for multiple models from the outset, with observability and governance as first-class requirements and a deliberate mix of open and proprietary technology. In creative AI that is harder than in text, because the routing decision cannot be reduced to language and token count. It has to account for input and output modality, visual style, aspect ratio, duration, character consistency, budget, speed, commercial purpose and what happens to the asset next.

Which is the whole point. Layer two is cheap and copyable. Layers one, three, five and six are where a decade of engineering habit shows up, and they are what we were actually buying in February.

Where FOTOhub's implementation goes further than a model picker

Adobe validated the direction. It did not implement the same product, and the differences run in FOTOhub's favour on four axes that we consider structural rather than cosmetic.

Scope. Premiere's picker routes video, inside a professional editing timeline. FOTOhub routes image, video, audio, chat, text, 3D and virtual try-on, across thirteen studios, on one credit balance — plus S3-compatible storage, no-code automation in Flows, a 314-endpoint REST API, SDKs in Python, TypeScript and PHP, an n8n node and a 28-tool MCP server that lets external agents call the whole thing. Orchestration that stops at one modality is a feature. Orchestration across seven is an operating system.

Automation. Adobe's user opens a list and chooses. That is a large improvement on five separate applications, and it is not the end state. Gabriel AI already classifies intent and selects the model, and FOTOhub's fallback path is not theoretical: writing his own analysis of the Adobe release, Mateusz Ulewicz noted that the illustration for it only rendered on the third attempt, because two Google models returned HTTP 403 and a different provider's engine finished the job. A founder publishing the failure of his own suppliers, in the post announcing his own validation, is the single most reassuring thing in this story. That is what a fallback layer is for, and it is why single-provider dependency is an operational risk and not merely a strategic one.

Vertical integration under the orchestrator. This is the axis no aggregator can copy with an integration budget. FOTOhub owns models beneath its own routing layer — IDA Q 1.0 for image, IDA Voice for speech, IDA Y for music and sound, plus the whole tooling tier — trained and served on its own GPU fleet in Frankfurt and Ireland. That puts a floor under gross margin, converts every future third-party price rise from an emergency into a routing decision, and is the only reason the platform can afford to give Professional and Business subscribers uncapped generation on the most popular image models. Adobe has the same insight with Firefly; almost no company at FOTOhub's stage has it at all.

Regulatory posture. A multi-model environment makes provenance harder, not easier: every element of a composite asset has a different origin. Since August 2, 2026, Article 50 of the EU AI Act has made labelling and transparency about AI use a legal obligation rather than good practice for anyone operating in the European Union. FOTOhub built for that from the start — its founder has been publishing on the AI Act since March 2026 — and a platform that has solved provenance internally saves that work for every business customer downstream. Adobe is pursuing the same objective through Content Credentials. Both are right; only one of them is a twelve-month-old company that treated it as a design constraint rather than a retrofit.

Dimension	Adobe Premiere + Firefly	FOTOhub.app
Entry point	Professional editing and post-production inside Creative Cloud	A native Creative AI environment for generation, automation and APIs
Modalities routed	Video and sound effects on the timeline	Image, video, audio, chat, text, 3D, virtual try-on
Engines	Firefly plus four named partner models	226 models from twelve providers, plus its own IDA family
Decision layer	The user picks a model in the context of the timeline	Gabriel AI classifies intent and selects or recommends; FOTOcore AI routes and manages resources
Failure handling	Provider-dependent	Health monitoring, model withdrawal from rotation, cross-provider fallback proven in production
Machine access	Within Adobe's application ecosystem	314 REST endpoints, three SDKs, CLI, n8n node, MCP server (both server and client)
Commercial model	Creative Cloud subscription plus generative credits	One subscription and one credit balance from $8 to $175/month, developer surface included at no surcharge
Core advantage	Deep integration with a mature professional workflow and enormous installed base	Breadth of modality, unified access, own models under the router, EU-first governance
Principal challenge	A consistent experience across partner models governed by different rules	Proving routing quality, reliability and scale against far better-capitalised competition

What this changes in how we underwrite AI companies

The uncomfortable half. Adobe's release is good news for our position and it also destroys a line of argument we have used for eighteen months.

Before September 8, orchestration was a thesis you had to sell. After it, orchestration is table stakes. The market has stopped asking whether a routing layer is necessary and started asking who builds a good one. That is a better question for a company with real engineering behind it and a worse question for anyone whose pitch deck was mostly a logo grid. The burden of proof has moved from the category to the execution — and, as FOTOhub's founder wrote himself, validation of a thesis by the largest player in the industry is good news; it is not an advantage. The advantage is whatever gets built on top of it.

Concretely, four things came off our diligence checklist and four went on.

No longer evidence of a moat	What we now weight instead
Number of integrated models	Routing quality measured on real production traffic, including fallback behaviour
Access to frontier APIs	Proprietary user context: libraries, brand assets, project memory, workflow history
Breadth of feature list	Workflow retention — do users come back next month, and does the interface stay legible as functions accumulate
Benchmark placement of any single model	Cost control and predictability: price shown before execution, per-operation attribution, own models anchoring the baseline load

On those criteria, our position reads as follows. FOTOhub converts 9.5% of users to paid, holds 62% week-eight retention and 145%+ net revenue retention, and has kept paid acquisition below 10% of signups while passing a million users — roughly $12 of capital per user against approximately $12.3 million raised in total, at a $180 million valuation in Q2 2026. Those are the numbers that tell us the routing layer is doing work, not the 226 in the catalogue.

What is still unproven, and we are not going to pretend otherwise

The catalogue is the copyable part. Anyone with an integration budget can add another API. If FOTOhub's advantage in twelve months is still "more models", we will have been wrong.
Adobe owns the professional timeline and will keep it. A frontal attack on post-production would be a mistake. The defensible ground is the set of workflows Adobe does not gather in one place: product photography, e-commerce assets, social content, fast marketing iteration, API-driven pipelines, and jobs that mix image, video, audio and 3D under one account.
Automated routing can fail badly. A router can pick a model that is cheap and inadequate, or excellent and non-compliant, or send data to a provider the customer never approved. Routing cannot be a black box optimising one metric; it has to weigh quality, cost, compliance, availability and informed consent, and it has to show its work.
More models can mean more chaos. Without strong architecture, a multi-model platform just relocates the mess from ten applications into one overloaded interface. Comparability across engines that treat the same reference image as inspiration in one case and a hard compositional constraint in another is an unsolved design problem industry-wide.
No SOC 2, no ISO 27001. The company states this plainly, and so do we. It is the next gate for the enterprise and public-sector contracts the platform is otherwise ready to serve, and it is on the agenda for the capital we are raising.

The next stage: the orchestrator chooses, and the author still decides

Today's multi-model workflow, in Premiere and in most of the market, still ends at a human opening a list. The next step is semi-automated routing: the platform reads task type, format, budget, deadline and organisational policy and returns two or three justified recommendations — strongest character consistency, lowest cost, fastest turnaround, cleared for commercial use, compliant with a given data policy — with the reason attached, not just the model name. After that, repeatable pipelines route themselves: cheap models for easy jobs, premium engines for hard ones, failed generations handed automatically to an alternative provider.

The line we care about, and the one we think most agentic products will get wrong: automation should eliminate technical decisions, not artistic ones. The system can recommend the engine, compute the cost, prepare the parameters and carry the context. Direction, quality criteria and final approval stay with the person. The best Creative AI OS will not be the one that generates everything without asking. It will be the one that knows when to decide, when to ask, and when to stop.

That is what DGE3 is funding. Our four bets for FOTOhub's second year, in descending order of conviction: more platform load carried by owned IDA models; agents that execute complete workflows rather than single tools; developer and enterprise revenue as a growing share of the mix; and distribution through protocol — MCP, plugins, marketplaces — rather than media spend.

DGE3 is assembling a $100 million Series A syndicate for FOTOhub: model training capacity, regional infrastructure, the developer platform, the two certifications above, and a United States go-to-market build. We are in conversation with growth funds, strategic investors and sovereign technology platforms. If that is your mandate, write to invest@vcdge3.com.

In 2024 the question was which model made the best image. In 2025, which made the best video. In 2026 it is a different question, and Adobe has now answered it in public: who builds the layer that holds all of these models together into one durable system of work. We placed that bet in February, on a company that had already placed it the previous September.

Frequently asked questions

What is AI model orchestration? The software layer above the models that decides which model runs a task and executes it in context. It requires six things: intent classification, a capability registry, routing with fallback, shared context, observability and workflow portability. A page of buttons wired to APIs is a model catalogue, not orchestration.

What did Adobe announce on September 8, 2026? Generative video and sound effects directly on the Premiere timeline, with a choice of engine — Firefly, Google Veo, Kling, Runway or Luma AI — reference frames taken from the current edit, the credit cost shown before generation, and the result delivered as an editable clip in place.

Did FOTOhub have orchestration before Adobe? Yes. FOTOhub launched on September 11, 2025 with multiple models behind one subscription, one credit balance and one library as its founding architecture — twelve months before Adobe's release. Its founder published the underlying fragmentation thesis on March 15, 2026, about six months ahead, and shipped the named FOTOcore AI and Gabriel AI layers during 2026.

What are Gabriel AI and FOTOcore AI? FOTOcore AI manages workflows, resources and routing across the catalogue. Gabriel AI sits above it, classifying intent, selecting or recommending the model, setting parameters and launching multi-step workflows across image, video, music, editing, chat and 3D. Gabriel runs on FOTOhub's own GPUs, which is why it can also be the free tier's chat model.

Is there a best AI model? Not usefully. An analysis of 80 models published in early September 2026 identified one image-quality leader in its dataset alongside eight distinct offerings on the price-to-quality frontier — top quality did not mean best economics. The right question is which model is good enough for this task at this budget, deadline, format, risk profile and legal requirement.

How many models does FOTOhub orchestrate? 226 as of September 2026: 63 image, 65 video, 30 audio operations, 35 chat, 20 text, plus 3D, virtual try-on and utilities, from twelve providers, including its own IDA Q 1.0, IDA Voice and IDA Y.

Why does a VC care about the orchestration layer? Because a leaderboard position turns over in weeks and a layer embedded in a customer's working day does not. Model count is not a moat; routing quality, workflow retention, proprietary context, cost control, compliance and distribution are.

Can I invest in FOTOhub? DGE3 Investment is assembling a $100 million Series A syndicate. Write to invest@vcdge3.com; company materials are at fotohub.app/investor-relations.