Construction · Deep dive

Arrakis

London- and Paris-based industrial-AI deployment platform from ex-Accel investor Rafael Quintanilla and three Palantir/Delivery Hero alumni — $38M raised in six months, $140M post-money at a July 2026 Series A, and a pitch to embed forward-deployed engineers inside aerospace, energy and manufacturing customers before Palantir, Cognite-inside-Schneider or SAP Joule Studio absorbs the category.

emerging

The question that decides it: Does an industrial-AI 'agent deployment platform' — services-heavy, forward-deployed, model-agnostic — compound into a durable layer above the frontier LLMs and the systems of record, or does Palantir AIP (already installed at Airbus, BP and most of Arrakis's target logos), Databricks Agent Bricks, Cohere North, and Schneider-owned Cognite absorb the industrial agent surface into their own SDKs and ontologies within 24 months — leaving Arrakis a very expensive Palantir tribute act that ends in an acqui-hire?

My take ★★ Very interesting ★ Interesting − Maybe ✗ Not interesting

Snapshot

Arrakis is what a former Accel investor and three engineers out of Palantir, Delivery Hero, Revolut, Datadog and ASML built when they concluded — in the founders’ framing — that the biggest AI returns will come from factories and supply chains rather than office software. Incorporated in London and Paris in January 2026, the company raised a $7.5M Accel-led seed in March 2026 and a $30M Blossom Capital-led Series A that closed at a $140M post-money valuation on July 22, 2026, per CEO Rafael Quintanilla to Fortune. The pitch is an “AI operating system” for industrial companies: a model-agnostic, sovereign platform plus forward-deployed engineers who embed at aerospace, energy, logistics, manufacturing, construction and telecom customers, deploy custom agents into existing systems in weeks, and tie half their fees to outcomes. Seven months in, no customer is named publicly, and the whole company is a bet that a Palantir-shaped services-plus-platform layer, unbundled and pointed at the physical economy, is a category — not a feature Palantir, Databricks, Cohere or Schneider absorb before Arrakis can build a moat.

Founding story

Quintanilla is the anchor. Educated at King’s College London and Imperial College Business School, he started full-time in college at pre-IPO Spanish fund platform Allfunds, then invested at Speedinvest and Paris seed firm New Wave before Accel hired him in London in 2022 to work early-stage AI, fintech and consumer, per his firm bio and Signal profile. Inside Accel he sat on the European AI thesis and helped US expansion for portfolio names like BeReal. By late 2025 he had — his cofounders and Accel’s own investment note say — concluded that the venture money pouring into horizontal LLM tooling and coding copilots was chasing the wrong tail: enterprise-grade AI would be won in industries where the pilot never reaches production, not in the seat-license SaaS categories where distribution is easy and value is small.

The engineering trio he pulled in tells the more specific story. Haroun Beltaifa and Romain Fouilland were both forward deployed engineers at Palantir out of Paris — Fouilland a Corps des Mines graduate, France’s elite technocratic track. FDEs are Palantir’s core sales weapon: engineers embedded at the customer who model the business as an ontology, wire up an agent into a live workflow, and only then hand off to a repeatable product. Mikhail Galkov’s CV — Delivery Hero, Revolut, Datadog, ASML — spans consumer-scale operations, fintech infrastructure, observability at scale, and semiconductor lithography, which is roughly the arc of what Arrakis says its platform is for. Accel, Quintanilla’s former employer, wrote the $7.5M seed in March 2026, an unusual but not unheard-of move. Blossom Capital led the $30M Series A four months later. Arrakis emerged from stealth on July 22, 2026, with a Fortune exclusive, coordinated coverage in Tech.eu, TechFundingNews, TheNextWeb, UKTN and Business Wire, and a $140M post-money — big for six months of trading and a headline that will price the next round.

How it works

The mechanics are — by design — recognizably Palantir. A customer engagement starts with Arrakis’s forward-deployed engineers embedded in the operational team: process interviews with plant supervisors, buyers, dispatchers or maintenance leads, mapped against the customer’s existing systems (ERPs, MES, SCADA, CMMS, EAM, TMS, procurement platforms). The applied-AI team tunes models on the customer’s own operational corpus — technical documents, historical work orders, purchase orders, sensor data, contracts. The platform then deploys custom agents into those existing systems rather than replacing them, on the deliberate premise (per company and Accel materials) that industrial customers cannot afford multi-year rip-and-replace migrations and that the value is in orchestrating what is already there.

Two design constraints matter. First, sovereignty: Arrakis says customer data and IP stay with the customer and are never used to train third-party models — the standard sovereign-AI pitch aimed at customers regulated by EU AI Act, defense contracting rules, or basic paranoia about frontier-lab data leakage. Second, model-agnosticism: the platform is built to swap between frontier LLMs (OpenAI, Anthropic, Mistral, open-weights), which is the counter to Cohere North’s tuned-in-house model and Palantir’s ontology gravity. The one disclosed customer proof point — a 90% reduction in procurement cycle times inside a NYSE-listed customer, achieved in weeks (company and multiple trade outlets, July 2026) — is what a procurement agent looks like when it can read RFQs out of email, query the ERP, negotiate with pre-approved suppliers via API, and route approvals: cycle time collapses because the humans are only asked when the agent’s guardrails flag exception. No customer is named. No second proof point is disclosed.

Product and business overview

The public surface is thin because the company is seven months old. What Arrakis describes is one platform with three moving parts: a model-agnostic, sovereign AI platform that hosts agents and connects them to enterprise systems; an applied AI team that tunes and evaluates models on the customer’s corpus; and a forward-deployed engineering organization that embeds at the customer, does the discovery, ships the first agent and stays for iteration. Sold across six industrial verticals (aerospace, energy, logistics, manufacturing, construction, telecom), with named early workflows in procurement, operational and financial-process visibility, and general “non-desk” work. Series A proceeds are earmarked for platform build, an expansion in the applied-AI team, and office openings in New York and the Middle East (Business Wire, July 22, 2026) — Middle East being the giveaway that the go-to-market includes sovereign-wealth-backed industrials (ADNOC, Saudi Aramco-class buyers) that pay Palantir-scale contracts and demand data residency.

Public job posts (Built In, Jobgether, fwddeploy.com, mid-2026) confirm the shape of the hire: “founding forward deployed engineer + equity” is the archetype, not “MLE” or “solutions architect.” The company is being built as a services-margin business with a platform on top, not the reverse.

Business model and pricing

No published pricing. The company describes revenue as fee-based with roughly half of fees tied to performance targets (Accel post, July 2026, and multiple trade write-ups). Read the way an underwriter would: the base fee is presumably the FDE loading (people-weeks against a platform license), plus a platform subscription; the variable fee is outcome-linked — savings realized, cycle-time reduction achieved, uptime improved. Outcome-based pricing sounds aligned but is operationally hard: attribution disputes with customers, revenue recognition drag, and cash-flow volatility. It also tells you the company does not yet have the pricing power to command a flat platform SKU, which is what an emerging category leader eventually needs to earn a software multiple. The comparable framework — Palantir’s ACVs in the mid-seven to eight figures with FDE loading front-loaded, moving to platform license as the deployment scales — is what Arrakis is implicitly targeting. The bear read is that a services-plus-outcome model at this stage sits closer to Big Four AI consulting than to a durable software business, which is exactly the C3.ai trap.

Traction over time

That is the entire disclosure at the time of this note. There is no ARR figure, no logo list, no headcount number, no retention data, no second workflow proof point. The $38M in six months and the $140M post-money are the numbers doing all the work. The scarcity is standard for a stealth reveal, and it is also the reason the openQuestion is what it is: the file to underwrite this company at $140M does not exist yet.

Market analysis

The addressable market Arrakis argues for is the industrial-labor line, not the software line. McKinsey’s widely cited range puts agentic AI value at $2.6-4.4T annually across business use cases; IDC projected $1.4T in global enterprise AI agent spend for 2027; Future Market Insights sized the narrower “industrial AI agents” market at $6.88B in 2026 with maintenance-and-reliability agents leading and edge/on-prem deployment dominant — the last two data points aligned exactly to Arrakis’s positioning. The structural forces behind the number are real: aging industrial workforces across Europe and North America, unfilled skilled-trades roles, a decade of failed digital-transformation programs that left industrial operators with data lakes and no agents, and post-2024 sovereignty pressure (EU AI Act, US export controls, defense contracting) that penalizes SaaS-in-the-cloud answers. The bull-case argument for a startup here is that this is exactly the market Palantir served for 15 years without competition because no other vendor could sustain the services-heavy motion; if frontier models make FDE-per-account cheaper, the services-plus-platform combination unlocks. The bear case is that the same forces are what got Schneider to pay $3.1B for Cognite, what pushed Cohere into North’s sovereign posture, and what is dragging Palantir’s revenue to a $7.65B guide — the money is not landing at the startups; it is landing at the incumbents that already own the account.

Competitive intel

Six directions of attack at once, ranged in frontmatter and expanded here. Palantir is the reference implementation and the direct threat: Foundry plus AIP plus the 2025 Agentic Foundry sells the identical forward-deployed motion to Airbus, BP, Ferrari and Kinder Morgan — logos on Arrakis’s target list — and Palantir posted $1.28B in Q1 2026 revenue on ~104% YoY growth per its 8-K. Arrakis’s honest pitch is not that it beats Palantir on capability today; it is that it is faster, sovereign by default, and does not lock the customer into Palantir’s ontology. Palantir’s counter is 20 years of proof and the fact that its FDEs come with an installed platform paid for. Databricks Agent Bricks attacks from below with a horizontal developer-first agent surface priced in DBUs on data the customer already stores in Databricks — 100,000+ agents built and a quadrillion tokens/year processed by the June 2026 Summit. Cohere North attacks laterally with the sovereign-tuned-model version of the same pitch and $240M+ ARR (TFN, 2026). Schneider-owned Cognite plus AVEVA attacks from the OT/factory-floor side; the June 2026 $3.1B deal is a bet that the industrial data layer wants to be owned by a hardware incumbent, not an app layer. SAP Joule Studio and IBM watsonx Orchestrate — both GA/expanded in May 2026 — attack from the ERP the customer already writes checks to. C3.ai is not so much a competitor as the cautionary tale: the same industrial-AI-platform pitch, ten years of runway, $246.7-250.7M FY2026 guide, and a market cap that halved in a year to ~$1.59B (Stock Analysis, August 2026). And underneath everything, LangChain and CrewAI commoditize the primitives, and Amazon’s June 2024 Adept acqui-hire is the market telling founders that horizontal agent-tooling exits go to the hyperscalers, not to standalone survivors.

History and evolution

• Jan 2026 — Arrakis Technologies incorporated in London and Paris by Rafael Quintanilla (ex-Accel), Haroun Beltaifa (ex-Palantir), Romain Fouilland (ex-Palantir) and Mikhail Galkov (ex-Delivery Hero, Revolut, Datadog, ASML).
• Mar 2026 — $7.5M seed round led by Accel (the firm Quintanilla had just left).
• Q2 2026 — Builds in stealth; first enterprise deployments across energy, logistics and industrial verticals; hires “founding FDE” cohort in London and Paris.
• Jul 22, 2026 — Emerges from stealth with a Fortune exclusive and a $30M Series A led by Blossom Capital, participation from Accel, GFC, MainObject and Rerail. Angels named: Olivier Pomel (Datadog CEO), Olivier Godement (OpenAI Head of Business Products), Junaid Hussain (Cambridge Aerospace, ex-Kingsway Capital). Post-money reported at $140M by CEO to Fortune. Company discloses 90% procurement-cycle reduction at an unnamed NYSE-listed customer; announces plans to open offices in New York and the Middle East.
• Late Jul 2026 — Coordinated coverage in Tech.eu, TechFundingNews, TheNextWeb, UKTN, Business Wire, Pulse 2.0, Euro-SD (defense trade), FinSMEs, Dealroom and StartupHub. Business Cloud pegs the round in sterling at ~£28M; UKTN at £22.5M.

That is the entire operating history at the time of this note. There is no launch flop, no pivot, no reorganization to record — the company is too young. The most notable “history” data point is negative: no customer name has been published, and the disclosed proof point is anonymized.

What people say

The case for. The investor signal is strong for the check size. Accel wrote the seed, wrote a public investment note framing this as the “last mile” of industrial AI, then followed on into the Series A. Blossom Capital — one of Europe’s more selective growth-stage funds — led the A four months later, at a valuation that implies belief the seed thesis is on. Datadog CEO Olivier Pomel and OpenAI’s head of business products Olivier Godement putting personal money in is unusual optionality: Pomel is a rare public-company operator who invests personally, and Godement is a Sam Altman-orbit executive whose angel checks are read by other operators. Trade coverage across Fortune, Tech.eu, TechFundingNews, TheNextWeb, UKTN and Euro-SD has been uniformly positive — the industrial AI category is investor-friendly right now, and the founder team ticks every box (ex-VC-turned-operator, ex-Palantir FDEs, dual EU footprint, sovereignty story that plays in Brussels and Riyadh alike). The 90%-procurement-cycle-cut proof point, if it holds up under diligence, is a strong hook — procurement is the workflow enterprises most want to automate and least trust vendors on.

The complaints. Every critical read of this company writes itself in the disclosure gaps. No customer is named — a company that has landed NYSE-listed enterprises inside six months usually gets at least one to press-release the deal, and Arrakis has none. No ARR, no headcount, no retention. The 90% number attaches to an anonymous customer and to a single workflow. The forward-deployed model is precisely the model industry critics — Anaplan’s CEO and ex-Palantir executive Manik Sharma among them (as documented in Forbes and Medium analyses of the FDE playbook, 2026) — argue creates customer lock-in but is a poor long-term platform strategy: the services line grows with revenue, gross margin stays around 60% rather than 80%, and the public markets discount the multiple. The founder-market fit critique is that this is a VC-founded services shop with ex-Palantir engineers, not a technical breakthrough — the durable moat that a $140M post-money implies is not yet visible. The category framing is the deepest critique: the industrial AI wrapper thesis has been tried at scale by C3.ai (repriced -49% in a year), and it is being attacked simultaneously by the two largest data platforms (Palantir, Databricks), the largest ERP vendors (SAP, IBM), the sovereign-model challenger (Cohere), and the OT-integrated automation giant (Schneider+Cognite+AVEVA). Every one of those has a distribution advantage into the exact CIOs Arrakis needs. And the tell nobody in the coverage flags: the $30M A closing four months after a $7.5M seed, at a $140M post-money, on no publicly disclosed revenue, is a valuation set by scarcity of European industrial-AI dealflow, not by numbers. When the next round prices, the numbers will need to exist.

Outlook: the open question

Arrakis works if, by roughly the end of 2027, three things are true. First, that at least two named marquee customers (a top-10 European aerospace or defense prime, a supermajor energy operator, or a top-5 European manufacturer) have taken Arrakis from procurement pilot into cross-workflow production — not just a procurement agent, but a second and third agent riding the same platform — and are willing to be quoted. Second, that revenue disclosure catches up to the valuation: a published $20-40M ARR run rate with net revenue retention above 130% would justify the $140M post-money at reasonable multiples; anything materially lower and the Series B either down-rounds or gets structured. Third, that the platform-versus-services mix moves visibly toward platform — declining FDE loading per new deployment, a repeatable industry-vertical template, and gross margin trending toward the low 70s. In that world, the openQuestion resolves in Arrakis’s favor: the industrial agent layer is a real category, the ex-Palantir team has unbundled Palantir’s motion faster than Palantir can respond, and Blossom’s bet compounds.

Arrakis fails if the systems already owning the account close the gap first. Palantir AIP plus the 2025 Agentic Foundry ships the same agent surface into Airbus, BP and Ferrari — no vendor swap required — before Arrakis can win them; Databricks Agent Bricks and Cohere North commoditize the sovereign, model-agnostic pitch at the mid-market; SAP Joule Studio and IBM watsonx Orchestrate bundle industrial agents into ERPs the customer already signs; Schneider-owned Cognite plus AVEVA colonizes the factory floor with an integrated OT-to-agent stack. In that world Arrakis becomes a very well-funded services shop with a modest platform, its 60%-margin outcome-based fees never mature into a software multiple, the FDE hiring plan gets expensive faster than revenue scales, and the endgame is an acqui-hire — to Palantir, to Databricks, to a hyperscaler, or (most on-brand for a French-founded industrial AI company) to Schneider or Dassault Systèmes as a bolt-on. The tells to watch over the next four quarters: whether a named customer of any size publicly attaches to the platform; whether Palantir or Cognite responds in the aerospace/energy accounts Arrakis pursues; whether the Series B is announced with a revenue number or without one; and whether the founding FDE cohort in London and Paris compounds — or churns after 18 months, as forward-deployed engineers characteristically do.

How a challenger would attack it

Attack the disclosure gap before the moat exists. Arrakis at seven months has no named customer, no ARR, one anonymized proof point, and a $140M post-money set by dealflow scarcity rather than numbers — which means its real asset is narrative momentum, and narrative is the cheapest thing to contest. A challenger (or, more realistically, an incumbent) simply publishes what Arrakis cannot: named industrial logos, audited outcome data, and a flat platform SKU. Palantir already does this into Airbus, BP and Ferrari with an installed data layer; Databricks prices agents in DBUs on data the customer already stores. A startup challenger would differentiate where Arrakis is structurally weak — the services mix. Its “founding FDE” hiring model means cost grows with revenue at ~60% gross margin; a rival that productizes one workflow (procurement agents, since Arrakis’s only proof point advertises the demand) as self-serve software on top of SAP and Coupa data undercuts the people-heavy motion at a fraction of the ACV, exactly the Agent Bricks play at the mid-market. The second vector is talent timing: forward-deployed engineers characteristically churn at ~18 months, and Arrakis’s equity-heavy pitch reprices badly if the Series B needs structure — poaching its FDE cohort in early 2028 buys the playbook without buying the company. Third, contest the sovereignty claim in its own market: Cohere North ships the sovereign pitch with a model tuned in-house, and an EU-flagged rival can note that a platform swapping between OpenAI and Anthropic APIs is thinner sovereignty than Brussels or Riyadh procurement actually demands.

Same playbook, new buyer

The unbundled-Palantir motion — FDEs plus a model-agnostic platform, half of fees outcome-linked — is a template, and Arrakis has claimed six verticals at once with maybe a few dozen people, which guarantees underserved lanes. The cleanest shift is down a size class: Palantir-style engagements and Arrakis’s implied seven-figure ACVs both ignore the Mittelstand — the thousands of €50M-€500M European manufacturers with the same aging-workforce and failed-digitalization pain but no budget for embedded engineering teams. A templated, partner-delivered version of the same procurement and maintenance agents, sold through the machine-tool and ERP channel those firms already trust, is a different business Arrakis’s FDE cost structure cannot chase while it hunts supermajors. The second shift is buyer type: Arrakis’s Middle East office points at sovereign industrials, but the defense-adjacent public sector — European armed forces logistics, port authorities, rail operators — has procurement rules that favor local, security-cleared vendors over a London-Paris startup with US-VC ownership; a nationally-flagged operator running the identical playbook wins there on compliance, not capability. Third, the outcome-fee model itself ports to asset-heavy services businesses — utilities O&M, facilities management, mining contractors — where “we get paid when cycle time drops” matches how the customer already buys, and where none of the six incumbents circling Arrakis’s aerospace-and-energy target list is paying attention.

Sources and further reading

• Exclusive: Startup Arrakis emerges from stealth with $38 million in funding to bring AI to industry (Fortune, July 22, 2026)
• Arrakis Comes out of Stealth With $38M to Help Industrial Enterprises Compete in the AI Era (Business Wire via Morningstar, July 22, 2026)
• OpenAI and Datadog leaders back AI deployment startup Arrakis (Tech.eu, July 22, 2026)
• Arrakis raises $38M from Blossom and Accel, with Datadog CEO and OpenAI exec joining the round (TechFundingNews, July 2026)
• Tackling Industrial AI’s Last Mile Problem: Our Investment in Arrakis (Accel, July 2026)
• AI startup Arrakis emerges from stealth with £22.5m round (UKTN, July 24, 2026)
• Ex-Accel investor raises £28m to build industrial AI company (BusinessCloud, July 2026)
• Arrakis emerges from stealth with $38 million to bring AI to factories and supply chains (The Next Web, July 2026)
• Schneider Electric announces agreement to acquire Cognite (Cognite/Schneider, June 2026)
• C3 AI Announces Fiscal Third Quarter 2026 Results (C3.ai investor relations, 2026)
• Palantir Technologies Q1 FY2026 8-K Press Release (SEC, 2026)
• Industrial AI Agents Market Size, Share & Forecast to 2036 (Future Market Insights, 2026)
• Palantir And Forward Deployed Engineering: What Should We Believe? (Forbes, July 10, 2026)