Are AI‑Native Services the New Software? Inside Sequoia’s Bet, GV’s Big Round, and the Rise of Recursive Superintelligence

What if the next trillion-dollar category isn’t another SaaS app—but a service that feels like a teammate? That’s the provocative idea Sequoia Capital partner Julien Bek is floating: services are the new software, and AI‑native firms will outcompete traditional SaaS by delivering intelligent, adaptive outcomes end to end. If that sounds like buzz, consider the money flows. As reported by Fortune, a four‑month‑old startup called Recursive Superintelligence—founded by alumni of Google DeepMind and OpenAI—has reportedly raised at least $500 million at a $4 billion valuation, in a round led by GV with participation from Nvidia, per the Financial Times. If you’re tracking the AGI race, that’s a blaring siren.

So, is Bek right? Are we watching software’s next mutation—one that looks less like tools and more like tailored, continuously learning services? In this deep dive, we’ll unpack the thesis, the signals behind the splashy Recursive round, and a playbook for founders, operators, and investors who don’t want to miss the turn.

Sources worth reading: – Fortune’s coverage of Bek’s thesis and the Recursive round: Fortune – Sequoia Capital: sequoiacap.com – GV (Google Ventures): gv.com – Nvidia: nvidia.com – OpenAI: openai.com – Google DeepMind: deepmind.google – Financial Times: ft.com

The Thesis: Why “Services Are the New Software”

Sequoia’s Julien Bek argues that generative AI and large language models (LLMs) aren’t just augmenting products—they’re enabling end‑to‑end services that feel bespoke, improve continuously, and deliver outcomes, not just features. In this framing, the buyer doesn’t pay for “seats” or “APIs” so much as for completed tasks, business KPIs, and guaranteed SLAs.

What “AI‑Native Service” Actually Means

Outcome‑oriented: You buy a result (e.g., “file the claim,” “ship the feature,” “close the books”), not a tool.
Full‑stack automation: From user interface to reasoning to integrations, the system does the work with minimal human glue.
Adaptive and learning: Performance gets better with data flywheels, fine‑tuning, and evaluation loops.
Human‑in‑the‑loop (HITL) by design: Experts oversee edge cases, teach the system, and provide auditability.
Moats from operations: Proprietary data, process graphs, and evaluation harnesses are as defensible as code.

In short: AI‑native services offer the convenience of a managed service with the scalability of software—because what used to be human labor in services is now largely inference and orchestration.

Why Static SaaS Struggles Against This Model

Traditional SaaS: – Offers tools that assume human operators – Ships features, not outcomes – Requires process change and training – Accrues value slowly as utilization climbs

AI‑native services: – Slipstream into existing workflows and APIs – Deliver “done” states with fewer handoffs – Learn from each transaction – Compete on business impact, not UI polish

When buyers are crushed for time and outcomes, the switch‑cost calculus changes fast. If an AI agent can handle 70–90% of the work with reliable escalation on the rest, the conversation becomes “Why not?”

Case in Point: Recursive Superintelligence’s Eye‑Popping Fundraise

Per Fortune’s report, Recursive Superintelligence—a startup just months old, founded by veterans of frontier labs—has raised at least $500 million at a $4 billion valuation, led by GV with participation from Nvidia, as reported by the Financial Times. The company is reportedly focused on advancing toward artificial general intelligence (AGI) through recursive self‑improvement—the idea that systems can iteratively enhance their own capabilities via evaluation and optimization loops.

Why it matters: – Talent premium: Teams from OpenAI and DeepMind carry enormous investor trust. Their playbooks for scaling models and systems are battle‑tested. – Capital intensity: If you believe AGI is within reach, compute is the scarce input. Strategic backing (e.g., from Nvidia) secures capacity. – Compressed timelines: Unicorn caps pre‑product used to be rare. In the AGI era, the market is placing big bets at idea‑maze stage.

Signal in the Noise: What This Round Tells Us

Services need serious R&D: Even if the near‑term product is a service layer, the frontier tech behind it is capital hungry.
Hardware‑software flywheel: Nvidia’s participation signals a well‑understood pattern—fund the apps that will saturate the chips.
Moats move upstream: If your service depends on cutting‑edge model capabilities (reasoning, tool use, planning), access to talent and compute is itself a moat.

Risks and Realities

Hype‑driven valuations: Big caps invite scrutiny. Can companies turn bleeding‑edge research into dependable services fast enough?
Regulatory heat: As AGI narratives intensify, expect policymakers to probe safety, labor impacts, and critical infrastructure dependencies.
Execution gap: World‑class research doesn’t automatically translate to delightful, reliable services. The last mile—evals, observability, and UX—decides winners.

How AI‑Native Services Create Defensible Moats

In classic SaaS, code and network effects drove defensibility. In AI‑native services, defensibility compounds across data, process, evaluation, and trust.

1) Data Flywheels and Proprietary Fine‑Tuning

Transactional data: Every task the agent completes (or fails) becomes training signal.
Private corpora: With consent and governance, customer data fine‑tunes models for domain‑specific excellence.
Synthetic data: Agents generate challenging cases to stress‑test themselves.
Personalization memory: Long‑lived memory architectures let agents adapt to org‑specific jargon, style, and preferences.

Result: A service that not only knows the job, but knows how your company does the job.

Useful primers: – LLM basics: Wikipedia: Large language model – AGI overview: Wikipedia: Artificial general intelligence

2) Workflow Integration and Compound Learning

Tool use and APIs: Agents execute steps in CRM, ERP, IDEs, ticketing systems—learning the happy path and exceptions.
Process graphs: Your workflows become explicit, machine‑navigable state machines that the system optimizes over time.
Operator feedback: Human reviewers label edge cases, close loops, and seed new capabilities.

The deeper the integration footprint, the stickier the service. Ripping out an agent that’s tuned to your processes is costly.

3) Trust, Evals, and Safety as Differentiators

Eval suites: Task success rates, hallucination benchmarks, regression guards—published and customer‑specific.
Governance: Data residency, PII minimization, audit trails, consent flows.
Risk controls: Harm/policy filters, role‑based access, reversible actions, sandboxed tool calls.

Trust isn’t a marketing slide—it’s a product surface. Services that make safety and observability first‑class win enterprise confidence.

Business Model Math: Can Services Have Software Margins?

The historic knock on services: lower gross margins and scaling limits. AI changes both—if automation climbs, margins start to resemble software.

Unit Economics to Watch

Task success rate (TSR): % of tasks completed without human escalation
Cost per action (CPA): Total inference + orchestration + human review per completed task
LLM COGS ratio: Model spend as a % of revenue
Human‑in‑the‑loop (HITL) ratio: Human minutes per task
Latency and throughput: Median and p95 time to completion, queuing behavior under load
Quality metrics: Accuracy, coverage, and customer‑defined KPIs

A target shape many teams chase: – TSR: 80–95% for well‑scoped tasks – HITL: <10% of cases, <2 min each – Gross margin: 70–85% with model routing, caching, and MoE architectures – SLA: p95 completion under customer‑acceptable thresholds

These aren’t guarantees; they’re north stars. But as model costs drop and orchestration improves, margins rise.

Pricing Models That Fit Agents

Per outcome: Pay per claim filed, ticket resolved, lead qualified
Value share: Revenue share or cost‑savings split
Subscription + meter: Base fee plus per‑task or per‑token usage
Performance tiers: Discounts for higher automation levels, premiums for faster SLAs or regulated workflows

Pro tip: Price how buyers get budgeted. Ops leaders often prefer outcome pricing; IT may prefer subscriptions with usage floors.

Capex and Compute Strategy

Strategic credits and partnerships: Work with hyperscalers and chipmakers to secure capacity.
Model routing: Use the cheapest capable model for each subtask; reserve frontier models for hard cases.
On‑prem/private options: For regulated industries, support VPC or on‑prem inference—even if at premium pricing.
Continuous cost engineering: Token efficiency, prompt distillation, low‑rank adaptation (LoRA), mixture‑of‑experts (MoE), and caching.

Compute is your new COGS line item—and your moat if you manage it better than competitors.

Go‑to‑Market Playbook for AI‑Native Services

Wedge Use Cases That Win Right Now

Look for tasks with: – Clear definitions of “done” – High volume, high variance – Painful handoffs and swivel‑chair work – Tolerant of partial automation with good escalation

Examples: – Research copilots that produce summaries, diligence, and literature reviews with sources – Coding agents that write tests, fix bugs, and refactor safely – Design assistants for assets, variants, and productionizing brand systems – L2/L3 customer support triage and resolution with knowledge grounding – Finance workflows: invoice matching, expense audits, close checklists – RevOps: lead routing, outreach personalization, quoting

Land with one job‑to‑be‑done. Prove TSR and ROI. Then expand.

Human‑in‑the‑Loop and the Automation Roadmap

Start human‑heavy: Use operators to guarantee quality and collect data.
Instrument everything: Log tool calls, error classes, resolution steps.
Automate class by class: Promote stable cases to auto mode; keep risky ones in review.
Publish your learning curve: Show customers TSR improvements over time. It builds trust and renewals.

Distribution: Sell Outcomes, Not Seats

Champions: Ops leaders who own KPIs (Support, Finance, RevOps, Eng Productivity)
Proof points: Before/after cycle times, cost per task, quality deltas, SLA adherence
Security reviews: Be pre‑emptive with data maps, DPA templates, and audit trails
Expansion motion: New workflows, deeper system integrations, higher automation tiers

Implications for Founders, Enterprises, and Investors

For Founders

Pick a market that rewards outcomes. Regulated or mission‑critical? Double‑down on evals and safety.
Build a serious evaluation harness. Make quality observable and improvable.
Get the data advantage early. With consent and governance, secure high‑signal corpora and feedback loops.
Design for enterprise trust from day one. Role‑based access, red‑team reports, incident playbooks.

For Enterprises

Pilot with guardrails. Start in a controlled workflow with clear SLAs and rollback plans.
Own your data contract. Ensure opt‑in/opt‑out controls, retention limits, and on‑prem options if needed.
Measure business impact. Insist on baselines and shared dashboards.
Train your teams. Pair operators with agents; codify learnings into SOPs.

For Investors

Underwrite to TSR and learning velocity. Can the team’s evals and data loops compound defensibility?
Watch for compute discipline. Smart model routing beats brute‑force burn.
Talent signal matters—but last‑mile excellence decides customers.
Favor outcome pricing with provable ROI and short time‑to‑value.

What Could Derail the “Services Are the New Software” Thesis?

Model performance plateaus: If reasoning, planning, and tool use stall, automation ceilings cap margins.
Regulatory brakes: Tighter rules on data use, model audits, and automated decisioning may slow rollouts in sensitive sectors.
Trust crises: A few high‑profile failures can reset buyer appetites and shift budget back to traditional tools.
Commodity creep: If everyone can fine‑tune similarly on comparable data, margins compress to the mean—unless your workflow and integration moats hold.

The Next 24 Months: What to Watch

Agent reliability benchmarks: Expect standardized, third‑party evals for common tasks, akin to industry SLAs.
Vertical specialists: Deep domain services in healthcare revenue cycle, insurance claims, legal ops, and defense logistics.
Hybrid models: Companies offering both SaaS tools and outcome‑priced services, letting customers slide along the automation curve.
Strategic chip alliances: More rounds with hardware co‑investors; compute reservations become part of due diligence.
Data governance features as table stakes: Differential privacy, redaction, granular retention, and customer‑controlled fine‑tuning.

Frequently Asked Questions

Q: What’s the difference between AI‑native services and SaaS with AI features? A: SaaS with AI features still assumes a human operating the tool. AI‑native services deliver outcomes—start to finish—with automation and HITL guardrails. You pay for results, not for using a feature.

Q: Can AI‑native services really hit software‑like margins? A: Yes, when automation rates are high, model routing is efficient, and human review is minimal. Many teams target 70–85% gross margins, though this varies by use case and regulatory overhead.

Q: How do these companies build moats if everyone uses the same frontier models? A: Moats come from proprietary data loops, workflow integrations, evaluation systems, safety/compliance posture, and distribution. The orchestration and operations stack can be more defensible than the base model.

Q: What pricing works best: per seat, per task, or value share? A: It depends on the buyer and workflow. Outcome‑based and value‑share align incentives but require robust measurement. Many teams blend a base subscription with per‑task or outcome pricing.

Q: How do enterprises manage risk with AI‑native services? A: Start with scoped pilots, require transparent evals and audit logs, enforce data governance and retention controls, and set SLAs with human escalation paths. Treat the service like a critical vendor.

Q: Will recursive self‑improvement (RSI) make these services obsolete or super‑charged? A: If RSI meaningfully improves reasoning and planning, services become more capable and cheaper to run—super‑charging the thesis. If progress stalls, services still deliver value but may cap at partial automation.

Q: Are talent exoduses from labs (OpenAI, DeepMind) a sustainable advantage? A: It’s a powerful near‑term signal for investors and recruits. Long‑term advantage depends on execution: building trustworthy services, distribution, and defensibility beyond pedigree.

The Clear Takeaway

Sequoia’s Julien Bek is tapping a real shift: buyers don’t want more tools; they want outcomes. AI‑native services—agents that do the work with human oversight—are poised to outcompete static SaaS by learning continuously, integrating deeply, and pricing to results. The reported mega‑round for Recursive Superintelligence—led by GV with Nvidia participating, per the Financial Times and covered by Fortune—shows just how strongly the market believes.

If you’re building, operating, or investing, optimize for the new stack: evaluation‑driven quality, data flywheels, trustworthy governance, and outcome‑based economics. Software ate the world. Services that feel like software—and deliver like great teams—are about to eat what’s left.

Discover more at InnoVirtuoso.com

I would love some feedback on my writing so if you have any, please don’t hesitate to leave a comment around here or in any platforms that is convenient for you.

For more on tech and other topics, explore InnoVirtuoso.com anytime. Subscribe to my newsletter and join our growing community—we’ll create something magical together. I promise, it’ll never be boring!

Stay updated with the latest news—subscribe to our newsletter today!

Thank you all—wishing you an amazing day ahead!

Are AI‑Native Services the New Software? Inside Sequoia’s Bet, GV’s Big Round, and the Rise of Recursive Superintelligence