Setsuna · Investor deck· 18 slides

Best viewed full-screen on desktop

SemSwitch, Inc. · Setsuna Investor deck · August 2026

Investor deck · 18 slides

The End of Warm Compute

Warm infrastructure exists because true cold creation was too slow. Setsuna changes that premise.

88 ms p50

16 GiB Firecracker microVM · request to verified ready · median of 1,000 true-cold launches

Use ← → or Space to navigate · F for full screen

SemSwitch, Inc. · Research preview

01 / 18

01 · The category thesis

Warm compute is a workaround for slow cold starts.

For decades, infrastructure paid an operational tax to make a machine look instant.

  • 01

    Snapshot the machine and restore it later.

  • 02

    Keep machines booted in a warm pool.

  • 03

    Preinstall, preboot and hold idle inventory.

  • 04

    Build a second cold path for misses, invalidation and failures.

02 · The hidden product

The workaround became an infrastructure product of its own.

Once creation is slow, teams must operate state artifacts, pool inventory and two lifecycle paths.

  • 01

    Capture, version, distribute and invalidate snapshots.

  • 02

    Secure guest-memory and machine-state artifacts.

  • 03

    Size pools for peaks; refill, health-check, drain and recycle them.

  • 04

    Handle stale state, compatibility and cold-path misses.

  • 05

    Debug behavior that differs between restored, pooled and fresh environments.

03 · Why agents break the workaround

Agents turn occasional cold starts into a core systems problem.

Thousands of short-lived, parallel, untrusted environments now sit inside interactive and training loops.

  • One user task can trigger dozens of execute-inspect-iterate steps.
  • Eval and RL rollouts fan out into thousands or millions of environments.
  • Burst demand makes pool sizing structurally inaccurate.
  • Generated code raises the value of a clean hardware-isolated boundary.

40%

of enterprise applications expected to include task-specific agents by end of 2026

Source · M1

$206.5B → $376.3B

AI-agent software spend, 2026 → 2027

Source · M2

$7.84B → $52.62B

AI-agents market, 2025 → 2030

Source · M4

Directional third-party market evidence for the agent category — not a Setsuna TAM.

04 · The new premise

Setsuna makes fresh creation faster than many restored or pooled service paths.

A 16-GiB Firecracker microVM reaches verified readiness without snapshotting or warm inventory — every launch, every time.

  1. 01New Firecracker process
  2. 02New KVM virtual machine
  3. 03Fresh guest kernel boot
  4. 04Verified networking, memory and workspace
  5. 05Destroy

88 ms

median request-to-verified-ready · 1,000 true-cold launches · 16 GiB guest

05 · Proof

The public record had no comparable documented true-cold result this fast.

Every published run — sequential true-cold launches per host and guest shape, timed from request to strict-ready. Raw samples and checksummed bundles are downloadable on the proof page.

HostGuestLaunchesp50p95p99maxRun
Azure Standard_D32ads_v732 vCPU · 64 GB1,000/1,000293 ms309 ms319 ms375 msAug 27, 2026
Azure Standard_F8amds_v74 vCPU · 32 GB1,000/1,000126 ms132 ms135 ms145 msAug 29, 2026
Azure Standard_F16ads_v78 vCPU · 32 GB1,000/1,000145 ms152 ms154 ms159 msAug 29, 2026
Azure Standard_F8amds_v7canonical2 vCPU · 16 GB1,000/1,00088 ms95 ms99 ms102 msAug 29, 2026
Azure Standard_E16ads_v72 vCPU · 16 GB1,000/1,00095 ms102 ms104 ms112 msAug 27, 2026
Azure Standard_E8ads_v72 vCPU · 16 GB1,000/1,00096 ms104 ms107 ms113 msAug 24, 2026
Azure Standard_D8ads_v72 vCPU · 16 GB1,000/1,00097 ms105 ms109 ms129 msAug 27, 2026
Azure Standard_F8amds_v74 vCPU · 16 GB1,000/1,00098 ms102 ms106 ms119 msAug 29, 2026
Azure Standard_F16alds_v78 vCPU · 16 GB1,000/1,000117 ms121 ms123 ms134 msAug 29, 2026
Azure Standard_E8ads_v62 vCPU · 16 GB1,000/1,000191 ms241 ms252 ms276 msAug 27, 2026

+ 12 more published shapes on semswitch.com/setsuna

Canonical run · Azure Standard_F8amds_v7

88 ms p50

p9999 ms

max102 ms

launches1,000/1,000

Managed sandbox create, independently measured — different clocks

  • E2B~717 msC15
  • Modal~2.4 sC15
  • Fly Machines~2+ sC14
  • Vercel Sandbox~1.9 sC15

Setsuna’s timer is host-local request-to-strict-ready. Public managed-service numbers may include network and control-plane spans. Never collapse the two into one ranking.

06 · Architecture deletion

From stateful workaround stack to create-run-destroy.

Setsuna can remove entire systems whose only job was hiding a slow fresh start.

Deleted for clean-slate jobs

  • Snapshot capture, versioning and restore.
  • Warm-pool sizing, refill, assignment and draining.
  • Cold-miss behavior divergence.
  • Persisted guest-memory artifacts in the normal path.

What remains

  • 01Create a fresh microVM.
  • 02Run the job.
  • 03Destroy it.
  • 04Keep images, persistent volumes and caches only where the workload actually needs them.

07 · Security + speed

Freshness without the historical latency tax.

Setsuna reduces persisted runtime state and reuse paths while retaining a Firecracker/KVM boundary.

  • Hardware-virtualization isolation.

  • Fresh kernel and guest userspace per execution.

  • No resumed memory/CPU/device state.

  • No pooled guest waiting for reassignment.

  • No customer state inherited from a prior execution.

08 · Where it wins

Make warm infrastructure optional — not universally forbidden.

Setsuna is strongest where the desired state is disposable and clean; snapshots remain valuable when continuation itself is the product.

Setsuna

Untrusted code execution, eval/RL rollouts, CI/test branches, one-shot tools, compliance-sensitive clean-slate jobs.

Snapshots / standby

Exact state continuation, long-lived coding sessions, forking from initialized state, heavyweight app/model initialization.

Hybrid

Persistent data outside the VM; fresh compute around each execution step.

09 · Economics and churn

The economic gain is idle capacity avoided plus engineering systems deleted.

Setsuna’s value compounds across compute, storage, operations and product velocity.

  • 01

    No idle warm guest fleet for clean-slate workloads.

  • 02

    No guest-memory snapshot storage and distribution in the normal path.

  • 03

    Fewer invalidation, compatibility and pool-health incidents.

  • 04

    One lifecycle path to test instead of warm-hit and cold-miss behavior.

  • 05

    Faster agent interactions and rollout completion.

10 · Category position

Setsuna is not another sandbox dashboard.

It is the runtime breakthrough that lets agent platforms simplify their own infrastructure.

  • 01

    Hosted API for developers who want the complete service.

  • 02

    Dedicated capacity for predictable performance.

  • 03

    BYOC for enterprise control and data boundaries.

  • 04

    OEM/runtime integration for existing agent and sandbox platforms.

11 · Go to market

Sell the deletion of warm infrastructure.

Start with teams that already maintain snapshots, pools or expensive environment setup and can measure the pain.

  1. 01Paid architecture and benchmark evaluation
  2. 02Integrate behind the customer’s existing sandbox interface
  3. 03Measure latency, failure paths, cost and code removed
  4. 04Convert to a hosted, dedicated or BYOC contract
  5. 05Publish customer-validated case studies

12 · Moat

Every qualified host and workload compounds the system.

Setsuna turns hardware-specific systems work into a growing profile and evidence asset.

  • 01

    Private per-host/per-shape runtime profiles.

  • 02

    Qualification registry and reproducible evidence.

  • 03

    Failure knowledge and production behavior across heterogeneous CPUs/clouds.

  • 04

    Customer workload contracts and integration depth.

  • 05

    Brand authority around measurable true-cold performance.

13 · The raise

$1.25M to make warm compute optional for production agent workloads.

Fund one senior systems hire, production control plane, paid pilots and customer-shaped qualification.

  1. 0118 months of runway.
  2. 023–5 paid design partners.
  3. 03Production API + BYOC pilot.
  4. 04Validated concurrency/reliability and unit economics.
  5. 05Next round raised on revenue and production workloads.

SemSwitch, Inc. is raising

$1.25M

Instrument
post-money SAFE
Runway
18 months
Primary contact
investors@semswitch.com

Appendix A · Use of funds

Illustrative 18-month allocation.

Large enough to stop fundraising again immediately, still inside the current pre-seed range for technical infrastructure.

  • Founder compensation + benefitsRemove survival pressure; preserve full-time founder focus.$190K
  • Senior systems / infrastructure engineerOne exceptional operator for runtime, productionization and reliability.$300K
  • Later technical capacity / specialist contractingSecond hire or targeted kernel/cloud/security help after first customer signal.$180K
  • Cloud, qualification and preview infrastructureNon-Spot qualification, production hosts, evidence automation and observability.$150K
  • Legal, accounting, insurance and securityCorporate hygiene, contracts, IP, SOC 2 preparation where demanded.$80K
  • Design partners, travel and toolsPilots, customer discovery, integrations and targeted GTM.$50K
  • ReserveCapacity shocks, delayed sales cycles and protection against under-raising.$300K
  • Total$1.25M

Target next-round evidence

  • Research preview converted into a reliable production API and restricted CLI.
  • 3–5 paid design partners, with at least one production deployment or BYOC pilot.
  • Production concurrency, reliability and cleanup contracts validated under customer-shaped workloads.
  • A credible unit-economics model at realistic utilization.
  • A repeatable qualification pipeline across the host families customers actually demand.
  • A commercial signal such as $250K+ contracted ARR or an equivalent committed pilot pipeline — a target, not a claim.

Appendix B · How Setsuna makes money

A focused entry wedge; four monetization paths in diligence.

Hypotheses to validate with buyers — not published price commitments.

  • Paid technical pilot

    Who pays
    Agent startups, labs, enterprise innovation teams
    Commercial shape
    $10K–$25K scoped evaluation
    Why it fits
    Fastest path to first revenue and proof of willingness to pay.

  • Hosted usage-based compute

    Who pays
    Agent platforms and developers
    Commercial shape
    Platform fee + per-second CPU/RAM/session usage
    Why it fits
    Validated by E2B, Daytona and Runloop pricing.

  • Dedicated non-Spot capacity

    Who pays
    Teams needing predictable capacity/performance
    Commercial shape
    Monthly minimum + active host-hour or committed capacity
    Why it fits
    Simple early operating model; customer receives a dedicated qualified host pool.

  • Enterprise BYOC

    Who pays
    Regulated or high-volume enterprises
    Commercial shape
    Annual platform license + support + usage/host fee
    Why it fits
    Customer pays the cloud bill; SemSwitch sells control plane, runtime and qualified profiles.

  • OEM / runtime licensing

    Who pays
    Agent platforms, clouds, CI/eval vendors
    Commercial shape
    Annual minimum + per-launch or per-host royalty
    Why it fits
    Setsuna becomes embedded infrastructure instead of competing for every end user.

Sources · 1/2

Research and evidence.

Source ids in slide footers map here. Market forecasts are directional signals, not a claim that Setsuna captures the forecast market.

  1. S1SemSwitch — “Anatomy of a True Cold Start”
  2. S2SemSwitch — Setsuna benchmark explorer and proof bundles
  3. S4SemSwitch — canonical public proof: Azure Standard_F8amds_v7, 2 vCPU · 16 GB, 1,000 consecutive true-cold launches
  4. T1Firecracker specification — ≤125 ms InstanceStart to /sbin/init (1 vCPU / 128 MiB)
  5. T2Firecracker project — KVM microVM isolation, minimal device model, jailer
  6. T3Firecracker snapshot support — guest memory + VMM/KVM state; operators secure snapshot files
  7. T4Firecracker snapshot versioning — memory dump, VMM/KVM/vCPU state, CPU compatibility constraints
  8. C1E2B infrastructure architecture — “a sandbox is a resumed snapshot”
  9. C2E2B pricing — Pro $150/month plus per-second CPU/RAM usage
  10. C3E2B Series A — $21M round, $32M total (vendor-reported)
  11. C4Daytona pricing — per-second billing; enterprise BYOC
  12. C5Daytona Series A — $24M
  13. C6Runloop pricing — Pro $250/month plus usage
  14. C7Runloop funding — $7M seed

SemSwitch, Inc. · Research preview

17 / 18

Sources · 2/2

Research and evidence.

  1. C8Blaxel seed — $7.3M
  2. C9Blaxel — 200–600 ms creation from template vs sub-25 ms standby resume
  3. C10Modal — unpacking sandbox startup latency (warm-pool guidance)
  4. C11Modal — sandbox pool example (size, TTL, readiness, claim logic)
  5. C12Modal — cold-start guidance (min_containers, buffer_containers)
  6. C13Modal — sandbox snapshot limitations
  7. C14Fly.io — suspend/resume: full VM state; ~2+ s cold start
  8. C15LogRocket — AI agent sandbox platform comparison, August 2026
  9. M1Gartner — 40% of enterprise apps to include task-specific agents by end of 2026
  10. M2Gartner — AI agent software spend $206.5B (2026), $376.3B (2027)
  11. M4MarketsandMarkets — AI agents market $7.84B (2025) to $52.62B (2030)
  12. F2Carta — State of Pre-Seed Q2 2026
  13. F3Heavybit — typical initial pre-seed check $500K–$3M
  14. E1Azure F8ads_v7 retail price tracker — $0.686/hour, South Central US, Aug 25 2026