GPU-accelerated ALM platform for insurance carriers
Valuations, reserves and full ALM projections in seconds to minutes on a workstation GPU. Nested stochastic VM-21 in minutes, not days. Your data formats, your network, your operating model — nothing to rip out.
Why the name InstaVal?
Every one of these is measured or shipping today. The rest of this page is the evidence.
The opportunity
Most carriers run an actuarial engine plus a warehouse, ETL, BI and reconciliation layer built around batch runtimes. InstaVal drops in beside all of it — nothing has to change on day one.
The machinery around the engine can shrink at your pace, as far as you want — the savings are real, but they are your call, on your timeline.
Coverage
Assets and liabilities projected together on one engine, with the statutory frameworks and strategy solvers built in.
Strategy solvers run root-finds over the full GPU projection engine: Defeasance delivers the market-consistent measurement that enables IFRS 17 Phase 2; Capital Funding solves funding IRR and required capital at a hurdle rate.
The core
Every life and health product runs on the same mathematical core: define the states a contract can be in and the transition probabilities between them, and the engine propagates the full state distribution every period. Term, whole life, UL / IUL / VUL, DI and LTC are one linear-algebra formulation — a new product is a new state model, not a new engine.
Exact distribution propagation, not point-estimate decrements — competing risks handled properly every period, by construction.
Matrix math is exactly what GPUs are built for — thousands of contracts' transitions propagate in parallel. The general formulation is the fast one.
Adoption
If your admin system can export a CSV, you're ready to run — most carriers load their first portfolio the same day. Conversion isn't a cliff, it's a gradual ramp: start with one product type, validate against your current system, expand.
Governance
The audit trail isn't a defended artifact. It's a reproducible computation — keep what's convenient, regenerate what's questioned.
Audit & compliance
Every run is identified, versioned and reproducible. Same inputs, same outputs — guaranteed.
Any single contract: per-period cash flows, state distributions, 3-basis reserves. In milliseconds.
Multi-sheet Excel export: policy features, state model, transition matrices, scheduled amounts, projection results, discount rates.
An auditor asks how you got a number — you rerun the exact calculation in front of them.
Sequences
A sequence is a versioned bundle of inputs — portfolio, assumptions, scenarios, parameters. Vary any input across a sequence and the platform runs the sweep: sensitivity analysis on anything, not just the shocks your vendor pre-built.
Shock any assumption, rate or parameter — no special project, no custom code. Define the sweep, click run.
Every leg has a sequence ID and versioned inputs. Reproduce any cell of the sensitivity grid, on demand.
Each leg runs in seconds to minutes, so a 20-leg sensitivity grid is an hour at most — not a weekend of batch scheduling. When a leg is cheap, you stop rationing questions.
Regulatory
US statutory modules are calibrated against published NAIC workbooks with documented error tolerances.
The numbers
Legacy engines run overnight — or over a weekend — and you review the results tomorrow. InstaVal runs in the same sitting: run, look, change an assumption, run again.
Scale — September 2026
Full ALM book, closed / production mode, 14 product classes (79% liabilities) — on an owned workstation with a ~$1K NVIDIA card, not a datacenter.
Cold vs. steady: the first scenario after boot is ~9 min (526 s, measured), then ~2 min each — the API server pays warm-up once. The current release pre-warms the request path: first answer on a cold macOS server went from 10.6 s to 2.5 s.
No GPU? The valuation / reserve API scales with cores — 27× at 32 cores, 46× at 64 on a rented x86 server, CPU only.
Deployment
Legacy: 3–5 days on a dedicated cluster, quarterly at best. InstaVal: about two minutes on one workstation GPU, on demand; about 25 seconds with model-point compression (10K→2K, 0.16% CTE error). Outer scenarios are independent — it scales near-linearly across GPUs.
Runs on hardware you already own or rent by the hour. No datacenter GPU. No appliance.
What is timed: the compute stage of a time-0 valuation — one projection or reserve pass per contract on a single basis, warm, excluding load and output. Not a multi-basis reserve trajectory. Measured Sep 1 2026 on both machines the same day; warm medians, items per second. Same engine, same outputs to the cent.
| Workload | Run type | Apple Metal · M1 Max (2022) | NVIDIA CUDA · RTX 4070 Ti | Faster |
|---|---|---|---|---|
| Bonds (10K) | 360-mo cash-flow projection, 1 scenario | 506K/s | 322K/s | Metal 1.6× |
| Mortgages (10K) | 360-mo projection, prepay + default | 212K/s | 147K/s | Metal 1.4× |
| MBS (5K) | 360-mo projection, prepay | 178K/s | 495K/s | CUDA 2.8× |
| Term life / Markov (2.8K) | Single-basis 120-mo state projection | 255K/s | 471K/s | CUDA 1.8× |
| Disability income / Markov (5K) | Single-basis 120-mo, 6-state Markov | 84K/s | 116K/s | CUDA 1.4× |
| Long-term care / Markov (5K) | Single-basis 120-mo, 7-state Markov | 80K/s | 100K/s | CUDA 1.2× |
| Fixed annuity CARVM (10K) | 30-yr greatest-PV-of-surrender reserve | 2.7M/s | 14.9M/s | CUDA 5.5× |
| Variable annuity (50K) | 360-mo fund projection × 10 scenarios | 209K/s | 214K/s | parity |
| ESG — 100 scenarios | Scenario generation, 360 mo | 17K/s | 33K/s | CUDA 1.9× |
| ESG — 1,000 scenarios | Scenario generation, 360 mo | 52K/s | 61K/s | CUDA 1.2× |
| ESG — 5,000 scenarios | Scenario generation, 360 mo | 62K/s | 28K/s | Metal 2.2× |
| VM-21 CTE70 | 51-scenario CTE70 reserve, 30 yr | 1.1M/s | 2.2M/s | CUDA 2.0× |
| Nested stochastic ASPA | 150 nodes: 50 outer × 3 dates × 50 inner | 139/s | 562/s | CUDA 4.0× |
| VM-20 UL PBR | NPR + DR + SR reserves, 70 yr | 57K/s | 150K/s | CUDA 2.6× |
| AG36 IUL | Implied rate + CRVM reserve, 20 yr | 18.7M/s | 18.9M/s | parity |
| Pensions (DB) · 100K participants | t=0 valuation · 3-basis projection + reserve trajectory | CPU, 8 threads, M1 Max: 0.92 s valuation · 2.16 s 3-basis trajectory | CPU only | |
Why this is hard to copy
Even if your current vendor got 100× faster tomorrow, every carrier's warehouse, ETL, BI and reconciliation stack would still exist — because those systems are built around the assumption that results are artifacts to be stored.
See it yourself
Sample portfolios for every product type are included — run your first valuation in the first five minutes, either way.
Next steps
We map your current infrastructure and identify where on-demand compute creates savings — and where it simply fits beside what you run.
Load your actual portfolio. Run your actual workloads. Measure actual wall-clock times.
A single workstation or a GPU VM in your own cloud. Parallel run alongside legacy until confidence is established.
The proof of concept is free. We're confident enough in the numbers to let the product speak for itself.
How licensing works
InstaVal runs on your own hardware — Mac (Apple Metal) or Linux (NVIDIA CUDA) — as a single compiled binary, air-gap capable. Usage is metered in credits: every valuation or projection consumes credits in proportion to what it computes (policies × scenarios × product complexity), debited before the run and refunded automatically if the engine errors. Nothing leaves your machine except an optional usage record.
Every tier runs the same meter, and a credit is one dollar — the number on your dashboard is the number on your invoice. Standard licences are prepaid: buy credits in the portal whenever you need more, no subscription, no per-seat licence. Enterprise licences are an annual contract with uncapped usage — run as much as you like, with no balance to watch — billed at the frequency you prefer; the meter keeps reporting consumption so the licence can be sized to your book. Terms are scoped in a conversation, not on a price list.
Hosted demo instances for prospects, and a UI-only download tier that needs no licence at all.
Metered per run, fractional credits, no ceiling and no minimum term on Standard.
Air-gapped installs replenish by pasting a signed token. Nothing phones home unless you let it.
A vendor-hosted cloud tier is on the roadmap and will be built when customer demand warrants it. Today, the same binary runs on a GPU VM you rent in your own cloud account — on-prem under the licence, with none of the idle cost.
Why us
InstaVal was built by someone who spent years on the wrong side of this problem — helping carriers navigate their data frustrations, watching actuarial teams wait hours for results they needed in minutes. Every design decision, from the GPU kernel architecture to the web UI, was made by someone who knows what it feels like to be the person waiting for the batch job to finish.
A 30-minute conversation: we walk you through the platform, run a live valuation against a portfolio shape that matches yours, and agree what a free proof of concept on your actual book would look like.