Applied Compute

1.1 Identity, headquarters, and business model

Applied Compute, Inc. is an enterprise AI company that builds what it calls "Specific Intelligence": proprietary agents trained on a customer's own data and workflows, deployed into production within the customer's own environment, and continuously improved through reinforcement learning on real usage traces. Multiple sources corroborate a 2025 founding by three former OpenAI researchers, with independent reporting (TechStartups, Grokipedia) narrowing this to May 2025. California Secretary of State filing records and a Comcast NBCUniversal LIFT Labs portfolio page both place the company's headquarters in San Francisco, at 251 Rhode Island Street #207; the regulatory filing further confirms the entity was originally incorporated in Delaware and registered as a foreign stock corporation in California on October 10, 2025, under document number B20250336266. Commercially, Applied Compute's own product materials describe a three-part platform - Train, Serve, and Improve - covering post-training of tool-using agents on customer data, low-latency production inference in the harness the model was trained in, and continuous online reinforcement learning from production feedback. The company markets its stack as "model-flexible," letting customers swap in newer frontier base models later without rebuilding their harness, data pipeline, or deployment stack, and states that its own engineers embed directly with customer teams rather than outsourcing development. These are company-sourced claims not independently re-verified against a customer's internal engineering process in this pass.[CO003, CO004, CO006, CO007, CO008, CO009]

1.2 Founders, leadership, and governance

Applied Compute was co-founded by Yash Patil, Rhythm Garg, and Linden Li, all Stanford University alumni who left technical roles at OpenAI to start the company. Patil was a key member of OpenAI's agentic Codex software-engineering effort and serves as Applied Compute's CEO; Garg was a core contributor to OpenAI's o1, the first reinforcement-learning-trained reasoning model; and Li worked on ML systems and infrastructure for reinforcement-learning training. Officer titles are not fully reconciled across sources: Applied Compute's California Secretary of State filing lists Rhythm Garg as Chief Financial Officer and Secretary, with Yash Patil as the sole listed Chief Executive Officer and registered agent, while Comcast NBCUniversal LIFT Labs' investor portfolio page instead lists Garg's title as Chief Technology Officer and Linden Li as Chief Architect. This is a genuine, unreconciled conflict in the sources reviewed rather than an editorial choice, and it is itself a governance-disclosure signal worth flagging: no independent, dated statement-of-officers document was located to settle it. The company's own account states that two-thirds of its team are former startup founders, including former top AI researchers and Math Olympiad winners, and Lux Capital's portfolio profile adds that team alumni include OpenAI reinforcement-learning-infrastructure veterans alongside people from Scale AI, Together, Two Sigma, and Watershed. No named executive beyond the three co-founders (e.g., a separate general counsel or head of sales) appears in any source reviewed, which is itself a key-person-dependence signal roughly a year into the company's life.[CO012, CO013, CO014, CO015, CO016, CO017]

1.3 Funding history, valuation, and investor base

Applied Compute's disclosed financing has moved in four public steps. Upstarts Media first reported, as an unannounced scoop, that the company had raised $20 million at a $100 million valuation in a round led by Benchmark partner Victor Lazarte, with Sequoia, Conviction, Hanabi Capital, Definition, and solo investor Zach Frankel also participating; a separate secondary write-up (StartupsUnion) instead dates the same round to "June 2024," a full year earlier than the 2025 founding corroborated elsewhere, an unreconciled and likely erroneous date that is reported here as-is rather than silently corrected. The Information then reported in approximately September 2025 that Applied Compute was in talks to raise at roughly a $500 million valuation, before the company emerged from stealth on October 29, 2025 with a publicly announced $80 million round from Benchmark, Sequoia, Lux Capital, Hanabi, Neo, Definition, Elad Gil, Victor Lazarte, and Omri Casspi; independent outlets placed the resulting valuation at roughly $700 million, a figure the company itself did not disclose at the time. By January 2026, The Information reported fresh talks at a $1.3 billion valuation, more than double the $500 million figure from three months earlier, which the company confirmed on April 8, 2026 with an $80 million Series B led by Kleiner Perkins (with Elad Gil, Lux, Greenoaks, Neo, and Hanabi continuing), bringing total disclosed funding to $160 million, per Applied Compute's own announcement and independently corroborated by TechCrunch citing PitchBook data. Latham & Watkins LLP served as Applied Compute's outside counsel on the April 2026 round. Taken together, the valuation moved roughly 13x in under a year, a pace of escalation that is itself a diligence flag independent of the company's underlying fundamentals.[CO018, CO019, CO020, CO021, CO022, CO023]

1.4 Scale, named customers, and cover-metric gaps

Applied Compute names three enterprise customers across its own case studies and funding announcements: Cognition (maker of the Devin AI software engineer and the Windsurf IDE), DoorDash, and Mercor. With Cognition, Applied Compute co-developed "SWE-check," a real-time bug-detection model embedded natively in Windsurf that Cognition's own blog describes as roughly 10x faster than the frontier model it replaced (Opus 4.6) while narrowing the accuracy gap to frontier performance on in-distribution evaluations. With DoorDash, Applied Compute built a calibrated automated grader and a reinforcement-learning-trained model to correct errors in AI-generated merchant menus during onboarding; an independent LLMOps case-study database (ZenML) corroborates that this delivered a 30% relative reduction in low-quality menus and was rolled out to all U.S. menu traffic, though the same independent writeup cautions that its account "is presented by Applied Compute, which naturally positions their tooling favorably," a promotional-bias caveat worth carrying forward. With Mercor, Applied Compute's custom-trained "Applied Compute: Small" model ranked #1 in the corporate-law category and 4th overall on Mercor's APEX-Agents leaderboard as of a February 2026 case study, ahead of Opus 4.5 and GPT-5.2, per company-hosted material not independently re-verified. Applied Compute's own fundraise post references working with "enterprises across the F500," implying customers beyond these three, but no additional names are disclosed anywhere in the sources reviewed. Headcount is not company-disclosed; RocketReach estimates 21-29 employees as of mid-2026. No source reviewed discloses revenue, ARR, or a customer count beyond the three named deployments.[CO030, CO031, CO032, CO033, CO034, CO035]

1.5 Milestone chronology and adverse signals

Applied Compute's chronology runs from its 2025 OpenAI departure and founding through a mid-2025 seed round, a September 2025 funding-talk leak, an October 10, 2025 California corporate filing, an October 29, 2025 stealth-exit financing, a January 2026 report of fresh unicorn-scale talks, a February 2026 Mercor benchmark result, an April 8, 2026 Kleiner Perkins-led Series B at $1.3 billion, a May 2026 Cognition product case study, and a July 2026 independent TechCrunch confirmation of unicorn status. Beyond the unreconciled officer-title conflict (Rhythm Garg as CFO/Secretary per the California filing versus CTO per Comcast LIFT Labs) and the unreconciled seed-round dating conflict (StartupsUnion's "June 2024" versus the 2025 founding corroborated elsewhere), no lawsuit, regulatory enforcement action, layoff, or security-incident disclosure naming Applied Compute was located in any source reviewed in this pass; this is recorded as an absence-of-evidence finding rather than a confirmed clean record, given how young and privately held the company is and how limited legal-database coverage was in this pass. The clearest adverse signal available from primary reporting is structural rather than event-based: a roughly 13x valuation escalation in under a year, a total lack of disclosed revenue or ARR against a $1.3 billion valuation, and governance disclosures (officer titles, headcount) that conflict or rely entirely on third-party estimators rather than company statements.[CO001, CO018, CO021, CO023, CO024, CO025]

1.6 Exhibits

2.1 Market boundary: what counts as enterprise AI agents

Applied Compute competes in enterprise AI agents and agentic software for internal business workflows, not the broader all-AI category. Its own material frames the opportunity as closing the "overhang" between a frontier model's raw capability and its realized utility inside one company's specific workflow, using two forward-deployed roles -- Forward Deployed Engineers (FDEs) who build evaluation harnesses and production environments, and Applied Research Engineers (AREs) who train and tune the underlying models. That places Applied Compute inside three overlapping spend categories: enterprise agent platforms and orchestration, reinforcement-learning post-training and evaluation-harness infrastructure, and forward-deployed customization labor. The same forward-deployed pattern -- pioneered by Palantir more than a decade ago and now productized as an AI agent inside Palantir Foundry -- is being copied at scale: AWS committed $1 billion in 2026 to a dedicated, agentic-first Forward Deployed Engineering unit serving named customers such as the NFL, the NBA, Cox Automotive, and Southwest Airlines. Excluded from this boundary are undifferentiated general-purpose model API consumption billed per token without an evaluation harness or RL customization layer, and incumbent workflow/RPA/CRM software that predates agentic AI; both are adjacent substitutes rather than core spend. Enterprises rarely choose only one path: 65% of surveyed technology leaders report a hybrid build-plus-buy architecture, with only about 10% relying on vendors alone, underscoring that internal build is a live substitute Applied Compute must displace deal by deal rather than a residual case.[CM001, CM002, CM003, CM004, CM005, CM006]

2.2 Market sizing: multiple lenses, no single reliable TAM

No single broad TAM figure is reliable for this market, so sizing has to stack several evidence-constrained lenses. Four independent market-research firms -- MarketsandMarkets, Grand View Research, Precedence Research, and Fortune Business Insights -- each size the standalone global "AI agents" software market within roughly a $7.3-8.5 billion band for 2025-2026, but disagree sharply on trajectory: MarketsandMarkets projects $52.62 billion by 2030 (46.3% CAGR), Grand View projects $182.9 billion by 2033 (49.6% CAGR), and Precedence projects roughly $294.66 billion by 2035 (43.57% CAGR). Layered above that is a much larger and differently scoped estimate: Gartner sizes 2026 spending on agentic-AI capability embedded across all enterprise software -- not just standalone agent vendors -- at $201.9 billion, roughly 25 times the standalone figures, because it counts embedded assistants and agent features inside existing applications rather than dedicated agent platforms. That gap is a measurement-boundary problem, not a factual contradiction, and it is compounded by Gartner's own total worldwide AI spending estimate of $2.52 trillion for 2026 (44% year-over-year growth, split roughly $1.37 trillion infrastructure, $452.5 billion software, $588.6 billion services), inside which the "agentic AI" sub-category alone is projected to compound at 119% CAGR from about $15 billion toward $753 billion by 2029. Applied Compute's own addressable slice -- enterprise agent platforms plus RL/post-training infrastructure plus forward-deployed customization -- sits somewhere inside the standalone-to-embedded range but has no independent third-party estimate isolating it; that gap is preserved as an explicit diligence ask rather than resolved with an invented number.[CM011, CM012, CM013, CM014, CM015, CM016]

2.3 Buyer, user, and payer segmentation

Applied Compute's four disclosed enterprise deployments span four distinct verticals, each sponsored by a domain-function leader rather than a single horizontal buyer persona: legal (Harvey, where a jointly post-trained model outperformed Opus 4.8 Max and GPT-5.5 xhigh on Harvey's 1,250-task Legal Agent Benchmark), software engineering (Cognition, where the SWE-check bug-detection agent runs roughly 10x faster than the frontier model it replaced), marketplace operations (DoorDash, where an RL-trained grader corrected AI-generated merchant menus), and labor-marketplace evaluation (Mercor, where Applied Compute's own small model ranks #1 in corporate law and 4th overall on Mercor's APEX-Agents leaderboard per a company-hosted, not independently re-verified, case study). Budget ownership for this kind of deployment is shifting away from central IT: a 2026 survey of 266 Fortune 50-Global 2000 technology leaders found line-of-business leaders are now the largest AI-tool buying group at 46%, matching or surpassing CIOs (38%) and CTOs (38%) for the first time. The typical adoption path still runs through a procurement and governance gate before budget commitment -- 84% of enterprises require security/compliance sign-off as non-negotiable and 70% want a self-serve sandbox trial first -- which favors vendors like Applied Compute that can embed engineers directly with a workflow owner rather than sell a shrink-wrapped product to central IT alone.[CM036, CM037, CM038, CM039, CM040, CM041]

2.4 Growth drivers and adoption constraints

Adoption is accelerating on several fronts at once. Gartner expects up to 40% of enterprise applications to carry task-specific agents by the end of 2026 (up from under 5% in 2025), McKinsey finds 88% of organizations now report regular AI use in at least one business function, and a 2026 CXO survey finds 42% of enterprises already run agentic AI in production (72% combined production-plus-pilot). The forward-deployed engineering model going mainstream at AWS and Palantir is itself a driver: it validates Applied Compute's FDE/ARE structure as an industry pattern, though it also signals that hyperscalers may compete directly for the same accounts. Constraints are equally concrete. Only one in five companies has a mature agent-governance model per Deloitte, and Gartner expects 40% of enterprises to demote or decommission autonomous agents by 2027 once governance gaps surface in production. Data readiness remains the top-cited blocker for a fifth consecutive year (58% of surveyed CXOs), and the return-on-investment bar is unforgiving: Gartner expects more than 40% of agentic AI projects to be canceled by 2027, and MIT-affiliated NANDA research found 95% of enterprise generative/agentic AI pilots fail to show measurable P&L impact, with vendor-partnered deployments succeeding roughly twice as often as purely internal builds. Layered under all of this is a capital-intensity constraint common to any RL-heavy vendor: the IEA projects data-center electricity consumption will nearly double to about 945 TWh by 2030, and Goldman Sachs projects global data-center power demand will rise 165% by 2030 versus 2023, a cost and supply backdrop that affects the price and availability of the compute Applied Compute's training and evaluation work depends on.[CM021, CM022, CM023, CM024, CM025, CM026]

2.5 Exhibits

3.1 Landscape: direct peers, platform incumbents, and build-your-own substitutes

Applied Compute does not compete for the same purchase decision as most of the companies enterprises mention in the same breath as "AI agents." Its own public materials frame the company as a reinforcement-learning post-training and evaluation-infrastructure partner: it trains proprietary, task-tuned models for named customers such as Harvey, Cognition, and DoorDash, and those customers then ship the resulting agent themselves under their own brand. That makes the practical competitive set unusually layered. Direct vertical peers -- Harvey (legal), Glean (enterprise search/agents), Sierra and Decagon (customer service), and Cognition/Devin (software engineering) -- each own a branded end-user agent product in a specific domain. Platform incumbents -- OpenAI's Frontier/AgentKit stack, Salesforce's Agentforce, and Microsoft's Copilot Studio -- bundle agent-building tools into existing developer platforms, CRM seats, or Microsoft 365 subscriptions enterprises already pay for. A public startup aggregator additionally lists dozens of smaller, earlier-stage entrants (LangChain, Agent Bricks, Manus AI, and others) circling the same broad space, though none carry the funding or named-customer evidence of the profiled direct peers. The most important nuance this chapter surfaces is coopetition: Harvey is simultaneously a disclosed Applied Compute collaboration customer for reinforcement-learning post-training of its legal agent, and an independently operating vertical incumbent that sells its own Assistant/Vault/Knowledge suite directly to law firms. The same forward-deployed-infrastructure-versus-branded-product distinction recurs with Cognition and DoorDash. Meanwhile, Sierra's own leadership describes building on a "constellation of" third-party foundation models plus its own fine-tuned proprietary layers -- an internal-build substitute path that, if it spreads, would reduce demand for an external RL-post-training specialist like Applied Compute even among companies that do not build a competing branded product.[CP001, CP002, CP003, CP004, CP008, CP010]

3.2 Competitor profiles: scale, funding, target customer, and strategic direction

Every named direct peer has disclosed a larger, more recent funding round than Applied Compute over the same 12-month window. Harvey raised $200M at an $11B valuation in March 2026, co-led by GIC and Sequoia, and reports more than 25,000 custom agents run by its law-firm and legal-department customers. Sierra raised $950M at a $15.8B valuation in May 2026 -- the largest of the group -- crossing $150M in ARR within eight quarters and counting more than 40% of the Fortune 50 as customers; founder Bret Taylor describes Sierra as "multiples larger" than its next-biggest rival by revenue. Decagon tripled its valuation to $4.5B in a $250M Series D in January 2026, just six months after a $131M Series C at $1.5B, and lists named customers spanning Avis, Hertz, Block, Affirm, Duolingo, and Oura. Cognition, maker of the Devin coding agent, raised $400M at a $10.2B valuation in September 2025 with disclosed ARR climbing from about $1M to $73M within a year, alongside reported layoffs and demanding working-hour expectations the same quarter. Glean raised a $150M Series F at a $7.2B valuation and positions itself as a horizontal build/deploy/orchestrate layer rather than one vertical wedge. The three platform incumbents compete on distribution rather than vertical depth. OpenAI now reports enterprise revenue above 40% of its total and is pushing a cross-system "Frontier" orchestration layer alongside AgentKit developer tooling, backed by "Frontier Alliances" partnerships with McKinsey, BCG, Accenture, and Capgemini. Salesforce's Agentforce reached roughly $800M in ARR in fiscal Q4 2026 (up 169% year-over-year) with more than 29,000 deals closed, though over 60% of bookings came from expansion inside its existing customer base rather than net-new wins. Microsoft's Copilot Studio has been adopted by more than 230,000 organizations to build custom agents, distributed as a low-code add-on inside Microsoft 365 and Teams subscriptions. Independent buyer guides note that even among the two closest customer-service peers, Decagon leans toward customer-owned engineering integration while Sierra leans toward a fully managed, forward-deployed delivery model -- a difference in delivery mechanism, not in underlying model technology.[CP007, CP006, CP012, CP011, CP016, CP017]

3.3 Capability, pricing, GTM/distribution, and trust posture compared

None of the profiled companies publish a self-serve list price for their core enterprise agent product, including Applied Compute itself. Harvey, Glean, and Sierra all sell custom enterprise contracts with undisclosed economics; independent 2026 buyer guides estimate Decagon and Sierra contracts typically start around $95,000-$150,000 per year and can reach $200,000-$350,000 once forward-deployed integration is included, though these are third-party estimates rather than vendor-published figures. Cognition's Devin is the one reviewed product with a disclosed pricing mechanism -- usage-based activity credit units -- which reviewers say creates a roughly $500-per-month cost floor and unpredictable total cost on long-running tasks. OpenAI is the most transparent on flagship API pricing but, like the others, keeps large-enterprise contract terms custom and undisclosed. Salesforce and Microsoft both fold agent pricing into existing CRM and Microsoft 365 subscription bundles, which makes apples-to-apples price comparison harder for buyers evaluating a standalone specialist against an incumbent add-on. On capability, the feature/capability matrix shows a consistent pattern: vertical peers (Harvey, Sierra, Decagon, Cognition) disclose strong domain specialization but leave their underlying model/RL-training depth largely undisclosed, while OpenAI discloses strong model depth but stays horizontal across verticals. Distribution and bundling power is concentrated in the three platform incumbents, which can attach agent tooling to renewals enterprises already budget for -- Salesforce's own disclosure that over 60% of Agentforce bookings come from expansion, not new logos, is direct evidence of that dynamic. On trust and regulatory posture specifically, none of the reviewed sources disclosed a detailed, comparable compliance/certification matrix across all nine entities, which this chapter treats as an open evidence gap rather than inferring a ranking from incomplete public materials.[CP036, CP035, CP034, CP031, CP021, CP027]

3.4 Switching cost, lock-in, multi-homing, and distribution power

Switching costs in this landscape come from three distinct mechanisms rather than one. First, forward-deployed or managed-engineering delivery -- used by Sierra, Decagon, and (per Company Overview) Applied Compute's own FDE/ARE model -- embeds vendor staff inside the customer's operational workflow, which raises the practical cost of switching beyond a simple software cutover. Second, data gravity: Harvey's Vault and Knowledge modules store and index a law firm's own documents and precedent inside Harvey's platform, so switching means re-indexing a firm's institutional knowledge, not just swapping a model endpoint. Third, bundling and distribution power: Salesforce and Microsoft can attach agent tooling to CRM and Microsoft 365 renewals their customers already sign every year, which is a structurally different and lower-friction sales motion than the net-new procurement decision every standalone specialist -- including Applied Compute -- must win. Multi-homing works differently at the model layer than at the application layer. Sierra's own leadership describes running a "constellation of" OpenAI and Anthropic models underneath its own fine-tuned layers, which means enterprise buyers of Sierra are not locked into one foundation-model vendor even though they are committed to Sierra's application layer. That same multi-homing option is exactly what makes internal build a credible substitute for some enterprises: a company with in-house ML talent can, in principle, replicate a lightweight version of any vertical peer's stack directly on OpenAI or Anthropic APIs. OpenAI's own Frontier pitch is built around this exact tension, explicitly positioning cross-system agents that "move across a company's systems and data" as more durable than agents "embedded within a single product," while Sierra's Bret Taylor frames Sierra's $950M raise as a defensive move to "invest aggressively" and hold its lead against a large number of well-funded rivals -- an implicit acknowledgment that even the reported category leader sees multi-homing and displacement as live risks.[CP037, CP038, CP039, CP040, CP041, CP042]

3.5 Moat durability, commoditization risk, and adverse competitor evidence

Applied Compute's core moat claim -- deep reinforcement-learning post-training and evaluation-infrastructure expertise -- faces a direct insourcing threat from two directions at once. Hyperscalers like OpenAI are pushing cross-system Frontier orchestration and their own post-training research budget, while richly funded vertical incumbents like Sierra openly describe building their own fine-tuned proprietary layers rather than buying that capability externally. Compounding this, Applied Compute's own proof points -- Harvey, Cognition, and DoorDash -- are themselves scaling toward or past multi-billion-dollar valuations, which gives exactly the customers used as evidence of product-market fit the balance-sheet capacity to insource RL post-training once the initial capability transfer is complete. The capital gap is stark: Sierra, Harvey, Decagon, and Cognition each raised single rounds of $200M-$950M at valuations 3x-12x larger than Applied Compute's own disclosed round within the same 12 months, giving every one of them more capital to bid for compute and research talent. Independent evidence adds further caution for the category as a whole, not just Applied Compute. Gartner forecasts that more than 40% of agentic-AI projects will be canceled by the end of 2027 on cost, value, or risk-control grounds -- a demand-side risk that compresses the addressable pipeline for every vendor profiled here. Independent 2026 testing of Cognition's Devin, one of the best-funded and most autonomous products reviewed, found real reliability gaps: roughly 78% success only on well-scoped bug fixes, unresolved security-vulnerability blind spots, and a documented cost floor near $500/month; Cognition itself laid off about 30 staff and offered buyouts to 200 more in August 2025 amid reports of demanding work-hour expectations, even as its valuation more than doubled the same quarter. Sierra's own founder publicly forecasts a market correction and a "culling effect" within roughly two years. Taken together, the read for Applied Compute is that its infrastructure-layer position avoids some head-to-head product competition, but it does not avoid the same capital-intensity, insourcing, and demand-durability risks that show up in its direct customers' and peers' own disclosures. One evidence gap remains genuinely open: no reviewed source clarifies whether DoorDash's parallel appearance as both an Applied Compute case study and an OpenAI enterprise customer reflects overlapping or entirely separate workloads.[CP043, CP044, CP045, CP030, CP031, CP046]

3.6 Exhibits

4.1 Revenue model and monetization

Applied Compute's own materials describe a three-stage "Agent Cloud" product -- Train, Serve, Improve -- delivered through an embedded, managed engagement model rather than a self-serve subscription. The company says its engineers "sit with" a customer's engineering team from the first evaluation through production deployment and ongoing online-RL improvement, a motion its own case studies illustrate concretely: engineers worked onsite at DoorDash's Sunnyvale office to translate production QA labels into an automated grader, and the Cognition engagement paired Applied Compute's RL stack with Cerebras inference to hit real-time latency targets for a production bug-detection agent. None of the three published case studies (Cognition, DoorDash, Mercor) discloses a contract value, minimum spend commitment, or usage-based fee schedule, and the company's own homepage, fundraise post, and launch essay describe the business only in qualitative terms. Benchmarked against OpenAI's, Anthropic's, Microsoft's, Cognition's own Devin, Together AI's, and Cerebras's published price lists -- all of which disclose at least a partial per-seat or per-token rate -- Applied Compute is the only vendor in this comparison set with a fully undisclosed pricing structure, consistent with a company still in an early, high-touch enterprise-sales stage rather than a metered-product stage.[CI001, CI002, CI003, CI004, CI005, CI006]

4.2 Unit economics and cost structure

Every classic unit-economics input -- gross margin, CAC, sales-cycle length, and net revenue retention -- is undisclosed for Applied Compute. The closest available comparator is Palantir, whose Foundry AIP delivery model likewise relies on forward-deployed engineers embedded with customers and which reported a GAAP gross margin of roughly 86.8% ($1.42B gross profit on $1.63B revenue) for the quarter ended March 31, 2026, showing that labor-intensive, embedded delivery is not inherently margin-dilutive at scale. Applied Compute's own cost stack plausibly combines GPU/inference compute (the Cognition case study names Cerebras as an inference partner), forward-deployed engineering labor (as in the DoorDash onsite engagement), and data/eval-harness costs (as with the Mercor partnership, where fewer than 1,000 expert-labeled tasks reportedly drove large evaluation-score gains, implying a comparatively low marginal data cost per engagement). An independent LLMOps review of the DoorDash case study explicitly flags that Applied Compute's own materials do not disclose "the total cost of development, ongoing inference costs, or quantified business impact," and cautions that the case study is promotional and should be read with appropriate skepticism -- a genuinely adverse, third-party check on the cost-benefit evidence the company publishes about itself.[CI015, CI016, CI017, CI018, CI019, CI020]

4.3 Capital adequacy and financing dependency

Applied Compute's financing history is unusually well documented for a private company: a roughly $100 million seed valuation in mid-2025, a $500 million funding-talk valuation reported by The Information in September 2025, an $80 million round at roughly $700 million disclosed at its October 2025 stealth exit, and an $80 million Series B at a $1.3 billion post-money valuation confirmed on April 8, 2026 by the company, its legal counsel Latham & Watkins, and Kleiner Perkins as lead investor -- bringing cumulative disclosed funding to $160 million, a figure TechCrunch's July 2026 unicorn roundup independently corroborates via PitchBook roughly three months after the round closed. What is missing is everything a lender or later-stage investor would need to assess capital adequacy on a forward basis: cash on hand, monthly burn, and runway are all undisclosed, and the company's only public statement on use of proceeds is qualitative ("grow the team, scale deployments, bring to market the first generation of agent workforces"). No source reviewed in this chapter discloses any venture debt, project-finance facility, or credit line; every round to date appears to have been priced as straight equity. Given the roughly 13x valuation escalation from seed to Series B in well under a year, the next raise looks more likely to be pulled forward by growth and investor demand than forced by an approaching cash shortfall, though that is an inference, not a disclosed trigger.[CI023, CI024, CI025, CI026, CI027, CI028]

4.4 Public disclosure gaps and financial verdict

A search of SEC EDGAR's company database for "Applied Compute" returns no filings tied to the 2025-founded startup -- the only similarly named registrant is an unrelated filer whose registration was revoked in 2006 -- confirming the company carries none of the standard public-company disclosures (revenue, margin, cash position, headcount) that are available for a forward-deployed-engineer comparator like Palantir through its 10-Q. Applied Compute's public case studies name exactly three customers, so the total paying-customer base, concentration risk, and renewal behavior all remain unknown. Category-level headwinds compound the disclosure gap: Gartner forecasts more than 40% of agentic AI projects will be canceled by the end of 2027 on cost and unclear-ROI grounds, McKinsey warns that a large majority of organizations have already encountered risky agentic behaviors that argue for new (uncosted) governance spend, and both the IEA and Goldman Sachs project sharply rising data-center power demand through 2030 that could raise the GPU/inference costs underlying any compute-heavy vendor's margin. Taken together, Applied Compute's financial profile today is best read as "well-financed but unverifiable": funding and valuation events are richly and consistently sourced, while every metric an underwriter would need to judge revenue quality, margin path, or capital runway is a confirmed, not merely under-researched, gap.[CI031, CI032, CI033, CI034, CI035, CI036]

4.5 Exhibits

5.1 Specific Intelligence is a managed specialization layer, not a general-purpose model API

Applied Compute defines the product in concrete workflow terms as a cloud for training, inference, and continuous improvement rather than as a standalone base model. The homepage breaks the platform into Train, Serve, and Improve: customers bring their own data, harnesses, and graders; Applied Compute post-trains tool-using agents across text, images, code, and structured data; and the resulting model is then served in production and updated from live feedback. The company repeatedly frames this as a specialization layer that sits above commoditizing frontier models, allowing the customer to keep its own reward functions, evals, and memory rather than outsourcing intelligence to a single upstream foundation-model vendor. That framing is consistent across the homepage, the fundraise post, the Modal customer story, and later technical research posts. Importantly, the platform is pitched as model-flexible rather than model-monolithic: Applied Compute says customers can train from the best available base model and later upgrade without rebuilding the harness, data pipeline, or deployment stack. In practice, that means the product promise is less “we have the best foundation model” and more “we help enterprises turn proprietary judgment into a durable, reusable training-and-deployment loop.”[CE001, CE002, CE003, CE004, CE011, CE037]

5.2 The technical architecture couples RL post-training with a context-and-memory system

Applied Compute’s public technical corpus shows two tightly linked product bets. The first is reinforcement-learning-based post-training in realistic, replayable environments. DoorDash’s deployment turned internal QA labels into an automated grader and then used that grader as a reward signal; Cognition’s SWE-check case recreated the production Windsurf harness during RL and then split post-training into capability and product-alignment phases; Mercor’s work used long-horizon RL over expert-labeled worlds and trajectory-level observability to catch reward-design failures. The second bet is that weights alone are insufficient, so agents also need a runtime memory layer. The Context Engine papers describe a Remember/Refine/Retrieve pipeline that ingests SaaS data and agent traces, distills them into a Contextbase, and exposes retrieval APIs to the agent at runtime; the follow-on “Memory in the wild” post shows the same loop pointed at Applied Compute’s own code traces, with critical-memory retrieval rising from under 10% to around 20% over two weeks. The company’s newer “neural cheat-sheets” work further suggests it wants context compression itself to become a trained capability, not just a prompting trick. Together, these materials imply a full-stack architecture: custom environments and graders produce better weights, while traces and context systems preserve what the model should remember between runs.[CE007, CE008, CE009, CE010, CE014, CE016]

5.3 Serving is optimized for long-horizon agents and controlled enterprise deployment

The Serve and Improve layers are as central to the product as the training loop. Applied Compute says customers can move from training to production with zero train-to-deployment mismatch, promote checkpoints instantly, and configure deployments around throughput, latency, and concurrency requirements ranging from asynchronous agents to real-time user experiences. The company also claims support for serverless deployment on its own cloud or full operation inside the customer’s VPC from a single control plane, plus single-tenant regional deployments and audit logging on every dispatch. Several research posts explain why these claims matter technically. The inference benchmark argues that agentic workloads differ materially from classic prompt/response serving because they create dozens of short generations, long-lived KV caches, and heavy-tailed tool latencies. The async-RL staleness post shows how utilization, queue capacity, rollout concurrency, and response-length tailness affect training quality once serving and learning are disaggregated. The public GitHub evidence is thin but directionally consistent: the organization exposes one Apache-licensed repo, `trie`, that replays inference traffic against an endpoint, matching the benchmarking narrative in the research posts. The result is a coherent picture of a platform optimized for trace-heavy, tool-using agents rather than single-turn chat throughput.[CE005, CE006, CE012, CE013, CE022, CE023]

5.4 Differentiation is credible, but technical risk still sits in verification, security, and productization depth

Applied Compute’s differentiation is strongest where it stays narrow and operational. Cognition’s SWE-check case supports the thesis that a smaller specialist can approach frontier accuracy on a single task with much lower latency and cost; Mercor’s benchmark work suggests the company can extract large gains from scarce expert data; and the Modal story indicates the team has already built around the infrastructure realities of rollouts, eval fan-out, and GPU-bound serving. Its 2026 publication cadence also signals a rapidly iterating research organization rather than a static services shop. That said, the public evidence has real limits. The best security claims — SOC 2 certification, data never leaving the customer perimeter, RBAC and audit logs — come from Applied Compute’s own marketing surfaces, with no public audit report, trust-center detail, or uptime history located in this pass. External guidance from Anthropic, Cohere, OpenAI, Microsoft, and CISA all converges on the same theme: production agents need tight orchestration, observability, and guardrails because flexibility increases cost and error surfaces. The independent Reward Hacking Benchmark adds a sharper warning for Applied Compute’s chosen paradigm, showing that RL-trained tool-using agents can exploit shortcuts unless environments are hardened. In short, the product story is unusually sophisticated for a very young company, but the public verification surface still lags the ambition of the architecture.[CE015, CE017, CE019, CE020, CE021, CE025]

5.5 Exhibits

6.1 Named roster shows real vertical breadth, but proof depth varies sharply by account

Applied Compute has progressed beyond anonymous logo slides. Its homepage names Cognition, DoorDash, Mercor, Harvey, Bridge, and Latch Bio, and the public case-study set fills in four distinct deployment archetypes: merchant onboarding at DoorDash, real-time bug detection inside Windsurf for Cognition, benchmark-driven professional-work specialization with Mercor, and legal-agent post-training with Harvey. That means the buyer/user/payer map is already diversified across product teams, engineering leads, operations teams, and domain-expert organizations. The common pattern is not a single horizontal seat sale; it is a specialized workflow inside a sophisticated customer that already has strong internal definitions of quality. The trade-off is that the public proof is uneven. DoorDash and Cognition look like genuine production deployments, Mercor and Harvey clearly look like high-value design partners, while Bridge and Latch Bio are presently best treated as testimonial-backed customer signals with limited deployment disclosure.[CU001, CU002, CU004, CU011, CU019, CU025]

6.2 DoorDash and Cognition provide the strongest production proof and the cleanest outcome metrics

DoorDash and Cognition are the most important references because both connect Applied Compute to live product surfaces rather than isolated benchmarking. DoorDash ties the work to merchant activation and menu accuracy, reporting roughly a 30% relative reduction in low-quality menus and a rollout to all U.S. menu traffic. DoorDash’s own June 2026 launch materials then show that AI-assisted onboarding has become part of a broader merchant-growth system, with merchants launching more than 35% faster and AI-powered websites converting at nearly 10% on average. Cognition provides similarly strong product evidence. Both Applied Compute and Cognition say SWE-Check is in production inside Windsurf and delivers 10x faster bug detection than the frontier alternative, while Cognition’s broader product updates show that Windsurf/Devin is becoming a multi-agent command center with meaningful distribution. These are the clearest signs that Applied Compute can attach to customer workflows where latency, precision, and iteration speed matter enough to justify a custom specialization layer.[CU004, CU007, CU008, CU009, CU010, CU012]

6.3 Expansion potential is credible, but public durability evidence is still proxy-based

The strongest expansion logic comes from the structure of the customer problems. DoorDash can apply the same merchant-quality infrastructure to onboarding, content, and ongoing merchant growth. Cognition can extend specialized models across local agents, cloud agents, review agents, and shared context surfaces. Harvey’s scale metrics suggest that a successful design-partner relationship could expose Applied Compute to a very large downstream usage base if benchmark gains feed live product modules. Mercor, Bridge, and Latch Bio each show adjacent verticals where customer-specific graders, expert data, or support judgments matter enough for post-training to be economically useful. But durability remains under-evidenced. Harvey’s 92% monthly adoption and 25+ hours saved are useful proxy metrics for the strength of the legal-AI end market, not direct proof of Applied Compute retention. Applied Compute discloses no NRR, GRR, churn, contract term, or share-of-wallet data, so the current public case for expansion is strategic and workflow-based rather than commercial and cohort-based.[CU010, CU016, CU018, CU022, CU028, CU029]

6.4 The largest open questions are concentration, procurement friction, and a services-heavy motion

The customer story is good enough to prove that Applied Compute can win sophisticated design partners, but not yet good enough to underwrite commercial durability. Most quantified outcomes come from company-authored case studies or customer-partner posts, not from independent customer-authored ROI material. The company does not disclose customer-count growth, revenue concentration, contract duration, renewal behavior, or whether its largest accounts are materially outsized. That matters because the public evidence also points to a services-intensive delivery model: Applied Compute says it embeds with customers, and Modal independently describes a deep engagement style around RL workloads. That can be a strength during the first wave of deployments, but it also raises scaling questions if every new account requires frontier-research attention. Independent adverse context reinforces the caution. CISA warns that agentic AI needs stronger oversight and Gartner’s 2026 service-ops agenda stresses that many pilots still need operating-model redesign and hard ROI proof before expansion. In practice, that means long sales cycles, slow security reviews, and a non-trivial risk that marquee logos overstate diversification.[CU003, CU034, CU035, CU036, CU037, CU038]

6.5 Exhibits

7.1 Competitive pressure is the top strategic risk because the control-plane story is crowding fast

Applied Compute’s core pitch is compelling: use a customer’s graders, data, and workflows to post-train a task-specific agent that outperforms generic models on that company’s exact use case. The problem is that the surrounding market is moving toward the same enterprise-agent control plane from several directions at once. AWS says Bedrock already serves more than 100,000 organizations and now offers AgentCore to operate agents securely at scale. OpenAI, Microsoft, Google, Salesforce, and Palantir all pitch secure enterprise agent deployment, governance, or digital labor on top of large existing distribution channels. Anthropic’s own agent guidance also undercuts any assumption that every workflow needs a heavyweight RL specialization stack; for many workflows, simpler orchestration patterns may be good enough. Applied Compute can still win where deep customization matters, but the burden of proof is now outcome-based: it must show that its specialist loop creates materially better economics or quality than increasingly capable platform defaults.[CR003, CR020, CR021, CR022, CR023, CR024]

7.2 Security and compliance burdens are real, and they can slow enterprise adoption before they cause headline incidents

The clearest non-competitive risk is not a known enforcement action today; it is the breadth of the governance burden around agentic AI. CISA warns that organizations need stronger oversight for agentic AI services. NIST’s AI RMF makes the diligence surface explicit: secure, resilient, accountable, transparent, privacy-enhanced, and bias-managed systems are the target standard. The EU AI Act adds a formal legal framework around AI risk, which matters because Applied Compute is selling into precisely the kinds of high-stakes enterprise workflows where audits, logging, explainability, and human-oversight design become procurement issues. McKinsey’s survey evidence that 80% of organizations have already seen risky agent behaviors reinforces the point that buyers are not just shopping for accuracy. They are underwriting operational trust. Applied Compute’s public trust claims are directionally good, but the absence of a rich public trust-center package means each regulated or security-sensitive deployment could still become a bespoke diligence exercise.[CR010, CR011, CR012, CR013, CR014, CR015]

7.3 The delivery model proves product value but creates people, partner, and concentration risk

Applied Compute’s customer wins show why the company exists: DoorDash, Cognition, Mercor, and Harvey all required customer-specific harnesses, graders, or benchmark environments. That depth makes the product hard to replicate quickly, but it also means deployments can stay dependent on scarce research talent for too long. The public record repeatedly describes an embedded, forward-deployed motion: researchers work with customer teams, DoorDash required onsite effort, Cognition used a replica of Windsurf, and Modal describes a deep integration model. On top of that, the company depends on a stack of outside providers and reference customers. It does not own the foundation model layer, it relies on cloud and execution partners, and its public go-to-market story is anchored by a small set of sophisticated flagship accounts. If any one of those accounts or partners weakens, the company could lose both revenue and social proof. The central question is whether Applied Compute can standardize enough of this motion before growth forces it to scale beyond a boutique forward-deployed team.[CR001, CR002, CR004, CR005, CR006, CR007]

7.4 Capital intensity is manageable today, but valuation expectations leave little room for sloppy execution

The April 2026 raise buys time, but it also hardens expectations. At a $1.3 billion post-money valuation and $160 million disclosed funding, Applied Compute is no longer judged like an early research boutique. It has to prove that the customization-heavy delivery model can compound into durable revenue before compute, energy, and security demands outgrow the available operating leverage. Gartner’s warning that more than 40% of agentic AI projects may be canceled by end-2027 should be read as a direct market-risk signal: even technically strong projects can fail commercially when cost, risk controls, or operating-model changes lag. Energy and data-center demand data from IEA and Goldman add a second layer of pressure by making compute-heavy post-training structurally expensive. The underwriting implication is clear: watch for whether public benchmark wins continue turning into repeatable production deployments, shorter implementation cycles, and clearer retention proof. If they do not, the valuation can unravel faster than the technical narrative.[CR009, CR016, CR017, CR018, CR019, CR032]

7.5 Exhibits

8.1 The financing fact is real, but the public record supports “track / research more” rather than a clean buy

Applied Compute clearly crossed the unicorn threshold: the company itself announced an $80 million financing at a $1.3 billion post-money valuation, and TechCrunch independently corroborated the same valuation and total funding to date. That validates the market mark. What it does not validate is whether $1.3 billion is already a fair entry point for a new investor. Public evidence supports high product ambition, credible flagship customers, and unusually strong case-study outcomes for such a young company. But it does not support a conviction price call because revenue, margin, retention, and customer concentration remain undisclosed. That makes valuation sensitivity the governing framework. At this price, even generous ARR multiple assumptions require meaningful scale that the public record does not yet show. The most honest recommendation from public evidence alone is therefore to track the company closely and move only if private diligence reveals stronger economics or price discipline improves. Public investors would normally demand this missing context before treating a financing headline as a durable valuation anchor.[CV001, CV002, CV003, CV013, CV025, CV026]

8.2 The thesis is strong customer-quality proof; the anti-thesis is evidence opacity at a premium price

The bullish side of the story is easy to see. Applied Compute has credible customer proof in DoorDash and Cognition, benchmark credibility with Mercor and Harvey, strong founder pedigree, and a backer narrative that fits where enterprise AI value may accrue. If those early deployments are the leading edge of a repeatable specialization platform, the current price may eventually look modest. The bearish side is more immediate. Almost all of the hard commercial questions remain unanswered in public: ARR, burn, gross margin, NRR, contract size, and round structure are all missing. ZenML’s skepticism about the DoorDash case-study economics captures the core problem well: quality of technical proof is not the same as quality of underwriting proof. Public evidence currently supports a high-quality company, not yet a fully supported $1.3 billion fair-value conclusion. Said differently: the market mark may be directionally right, but the evidence base still looks thin for precision.[CV004, CV005, CV006, CV008, CV015, CV016]

8.3 Comparable context suggests premium AI valuations are possible, but Applied Compute has not yet matched the public disclosure behind the best comps

The most useful private comp in the current file is Harvey. CNBC reports Harvey at an $11 billion valuation with about $190 million of ARR as of January 2026, which implies that the private market will award extraordinary multiples to enterprise-AI companies with visible scale, strong adoption, and elite investor support. That helps prove that the market can price premium application-layer AI expensively. It does not prove Applied Compute deserves the same treatment today. Harvey pairs valuation with disclosed ARR and visible downstream user scale, while Applied Compute has only a handful of recent public deployments and no disclosed economics. Public-company anchors like Palantir, Salesforce, Microsoft, and Alphabet are still useful, but mostly as disclosure and scale comparators rather than clean multiple comps. The resulting scenario frame is asymmetric: bull-case upside exists if private diligence reveals far stronger ARR and repeatability than the public record implies, but the base case remains below the current mark unless those private metrics are unusually strong. That is why scenario discipline matters more here than nominal enthusiasm about the category.[CV010, CV011, CV017, CV018, CV030, CV034]

8.4 The decisive question is whether private diligence can convert narrative strength into economic proof

This investment does not fail first on product quality. It fails if marquee case studies do not convert into repeatable economics before the market becomes even more competitive and transparent. That is why final diligence matters so much. Investors need current ARR, gross margin, NRR, customer concentration, and the exact structure of the April 2026 round before underwriting the price. Without those, governance opacity and commercialization risk deserve a discount. Gartner’s agentic-AI cancellation warning and the abundance of competitor pricing transparency further increase the cost of wishful thinking. If diligence can demonstrate that Applied Compute already sits near the ARR and retention thresholds implied by premium private AI multiples, the current mark can be defended. If not, the company may still be excellent while the entry price remains rich. That distinction is the core of the recommendation. The burden is on the company to convert technical prestige into transparent, repeatable, and durable economics.[CV014, CV019, CV020, CV021, CV022, CV023]

8.5 Exhibits