Gestell
Gestell performs compiled GPU execution analysis: it statically reads emitted PTX and SASS and returns a compiled-output diff for inference-engine maintainers, platform teams and enterprise architects. A second track covers legacy mainframe harmonization. Access starts with direct contact.
What is Gestell?
Gestell is an engineering and research company rather than a self-service product. Its public site is small: a positioning page, an article index, Terms of Service and a privacy policy, eight URLs in total.
The stated focus since 2026 is compiled GPU execution analysis. The company describes itself as building tools to understand and advance the frontier of GPU performance, and its own meta description says it studies PTX, SASS, compiler lowering, and GPU execution behavior. The central deliverable is a compiled-output diff: which kernel families appeared, disappeared or changed between two states of an inference engine, and which instruction-level execution structures moved.
The published methodology is specific. Compiled artifacts are collected across comparable serving configurations, the emitted PTX and SASS are read statically, kernels are grouped into families, presence is compared, instruction deltas are computed, and configured review policies are evaluated. Those policies are declarative and versioned: an identifier, an effect, matching criteria such as a kernel glob, an operation or a PTX layer, exclusions, a trigger threshold and a human-readable reason. One documented example, review-gemm-when-tma-expected, flags a GEMM kernel that was expected to use TMA but emits repeated manual global-to-shared copies instead. The scope is as notable for what it excludes: the analyzer does not execute kernels, does not run the model, collects no profiler counters and measures no runtime. Three published SGLang case studies illustrate the approach: pull request #26588, where two mathematically equivalent fused paths were reverted after a Gemma4 GSM8K regression; DeepGEMM kernels turned into a CI gate; and an Ampere-versus-Hopper execution comparison.
A second track is legacy system harmonization, which Gestell defines as the process by which the new is folded into the old, accelerating both. The claim there is that AI agents can ingest and structure millions of lines of legacy code, surface business logic buried in monthly batch jobs, map dependencies that were never documented and identify dead code that still costs licenses. A public repository, Gestell-AI/zowe-mcp, supports that track as an MCP server for the Zowe CLI.
Note that the Terms and the privacy policy govern something they call a Data Ingest Platform, and that a web application sits behind authentication at platform.gestell.ai with no open sign-up. There is no pricing page, no product page and no public API documentation; the entry point is email.
What it does
- Produce a compiled-output diff showing which kernel families appeared, disappeared or changed between two states of an engine
- Statically read the emitted PTX and SASS without executing a single kernel
- Group kernels by family and compare their presence from one state to the next
- Measure instruction-level deltas between states
- Apply configured review policies and raise a flag when a threshold is reached
- Collect compiled artifacts across comparable serving configurations
- Ingest and structure millions of lines of legacy code, map undocumented dependencies and identify dead code
When to use Gestell / When not to
A quick filter to help you decide if Gestell is the right fit.
When to use Gestell
- Teams maintaining a GPU inference engine who need to see what compiler lowering actually changes
- Reviewers of pull requests where a mathematically equivalent source change can still shift floating-point execution
- Platform engineers setting up a CI gate on compiled output, evaluated on every commit
- Engineers migrating across GPU architectures, such as the documented Ampere-to-Hopper case, who need to know which kernel families moved
- Enterprise architects mapping an undocumented mainframe estate before transforming it
When not to use Gestell
- Anyone looking for a benchmark or a ranking of inference backends, which Gestell explicitly rules out
- Teams that need a runtime profiler: the analyzer executes no kernels and collects no profiler counters
- Buyers who expect to sign up alone and try the product, since there is no self-service path and no published price
- Small teams with no low-level GPU stack and no legacy estate to harmonize
- Organizations shopping for a COBOL-to-Java translator, an approach the company itself rejects
How to use Gestell
A typical end-to-end flow, from setup to results.
- Read the published articles first to check whether compiled-output analysis matches the problem you have
- Write to hello@gestell.ai: the Contact link in the site navigation opens a mailto to that address, and there is no public sign-up form
- Ask for the full compiled-output analysis reports, which the SGLang article says are available on request
- Scope the engagement directly with the team, since no self-service plan and no published price exist
- Once you are a customer, sign in to the web application at platform.gestell.ai; the Terms make you responsible for keeping your credentials confidential
- Prepare comparable serving configurations whose compiled artifacts can be collected
- Declare your review policies as versioned rules with matching criteria, thresholds and human-readable reasons
- Wire those rules in as a CI gate evaluated on every commit
- Read the compiled-output diff to see which kernel families and which instruction-level structures moved
- For mainframe work, install the open-source zowe-mcp server from GitHub, independently of any commercial engagement
Pros & Cons
Pros
- Unusually honest scope: the company states plainly what its analyzer does not do, from kernel execution to profiler counters and benchmarking
- The methodology is published in detail and broadly reproducible in outline
- Case studies rest on public, identifiable pull requests and projects, including SGLang PR #26588, DeepGEMM and FlashInfer
- Addresses a real blind spot: the divergence between algorithmic surface and compiled surface, which source review, benchmarks and profilers all miss
- The output is designed as a compact structured artifact, usable in CI and readable by a maintainer or an agent instead of raw SASS
- Review policies are configurable rather than a fixed rule set
- A public, freely installable MCP server and a senior advisor with a verifiable mainframe-modernization record at Micro Focus, Microsoft, DXC and AWS give the claims some external footing
Cons
- Very thin public surface: no product page, no pricing page and no API documentation, with eight URLs in the entire sitemap
- No price, no plan and no trial, so cost cannot be assessed before a sales conversation
- No screenshot, no demo and no product walkthrough of any kind
- Brief and generic legal documentation: no GDPR mention, no data processing agreement, no subprocessor list, no retention period
- No postal address, no named legal entity and no identified governing jurisdiction in the Terms
- The headline and the contractual documents diverge: the home page sells GPU analysis while the Terms and privacy policy govern a Data Ingest Platform
- No executive is publicly named, only a senior advisor, and nothing is published about where customer data is hosted
Pricing & Plans
No pricing information is published on the Gestell website. No free plan is announced, no entry price is stated and no billing model is given; the pricing URL returns a 404 and the complete sitemap contains no commercial page. Cost is therefore established through direct contact with the company.
Data, GDPR & hosting
A consolidated view of how Gestell handles your data.
GDPR overview
The GDPR is never named on the Gestell site: the term appears in none of the pages reviewed, including the Terms and the privacy policy. What the policy does provide is a conditional list of rights, introduced by Depending on your location, you may have the following rights regarding your personal data, and covering access, rectification, erasure, restriction of processing, portability and objection. Requests go to hello@gestell.ai. Beyond that the documentation is silent: no legal bases for processing, no Article 27 EU representative, no data protection officer, no data processing agreement offered or mentioned, no named subprocessors, no retention periods, no reference to international transfers or standard contractual clauses. The Terms add to the uncertainty by deferring to the jurisdiction in which Gestell is established without ever naming it. European buyers should treat GDPR alignment as something to negotiate, not as something documented.
Who owns the data?
The Terms of Service, effective 17 February 2026, separate two sets of rights. Anything you upload, submit or provide through the Service counts as User Content: you remain solely responsible for it and for the consequences of submitting it, and you warrant that you hold the necessary rights and infringe no third party. The Terms do not transfer ownership of that content to Gestell, but they grant the company a worldwide, non-exclusive, royalty-free license to use, reproduce, modify and distribute it in connection with the Service. In the other direction, the content, features and software of the Service belong to Gestell or its licensors. No named third-party recipient is listed.
Reuse rights
The privacy policy, effective 17 February 2026, lists four categories of collected data: identification details (name, email, contact information), payment data (card and bank details, transactions), usage data (access times, pages viewed) and technical data (IP address, browser, operating system). Five purposes are stated: providing and maintaining the platform, improving the experience and its features, communicating with users including updates and support, monitoring usage and performance, and complying with legal obligations. On sharing, Gestell states that it does not sell, trade or otherwise transfer personal data to outside parties, with two exceptions: trusted service providers that help operate the platform, and legal requirements or the protection of rights. No provider is named and no subprocessor list is published. Model training appears nowhere in the document: it is neither authorized nor ruled out, so a user who cares about that point has to settle it in the contract rather than in the policy.
Data retention & training
Things to keep in mind
Risks and trade-offs to weigh before adopting Gestell.
- The Terms are governed by the jurisdiction in which Gestell is established, a jurisdiction the document never names, which leaves you without a known forum in a dispute
- No registered legal entity and no legal suffix appear anywhere: Gestell is the only name given, in the contact section of the Terms, and no postal address is published on the site
- The license granted on User Content is broad, being worldwide, non-exclusive and royalty-free, with the right to modify and distribute
- No data retention period is stated, and no commitment is made about the country where data is hosted
- The privacy policy neither states nor denies that customer data is used to train models, so the question stays open unless you settle it contractually
- A third-party B2B visitor-identification script, RB2B, loads on the site without appearing in the privacy policy
- Gestell may suspend or terminate access without prior notice or liability, and its liability is capped at the amounts paid over the last twelve months
Setup & Integrations
Technical difficulty
High, and front-loaded. There is no self-service installation: access begins with an email exchange, then a login at platform.gestell.ai. On the GPU side, the work assumes you can produce compiled artifacts across comparable serving configurations, which means real command of your own inference stack. Review policies are written as declarative rules, with an identifier, matching criteria, a threshold and a reason, then wired into CI. The material examined is PTX and SASS, so reading the results calls for low-level GPU literacy. The zowe-mcp server installs independently from GitHub. No public setup or onboarding documentation exists.
Deployment
Integrations
Behind Gestell
Social
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
What does Gestell actually do?
Is Gestell a profiler or a benchmark suite?
What is a compiled-output diff useful for?
Can it be used in continuous integration?
Which case studies has Gestell published?
How much does Gestell cost?
Can I try it on my own?
How do I get in touch?
Does the site say anything about the GDPR?
Does Gestell also work on mainframes?
Should you pick Gestell?
Gestell is a niche tool aimed at teams working at the level of compiled code. Its proposition, making visible what compiler lowering actually changes between two states of an inference engine, answers a real and documented problem: a source change can be mathematically equivalent and still move floating-point execution, and neither source review, nor benchmarks, nor runtime profilers will show it. The published SGLang work, including the reverted pull request #26588 and the DeepGEMM CI gate, is today the best evidence of the team's know-how.
Commercial maturity is another matter. There is no product to try, no price to compare and no legal entity named anywhere. The Terms defer to an unnamed jurisdiction, the privacy policy never mentions the GDPR, and the contractual documents govern a Data Ingest Platform while the home page sells GPU analysis, a mismatch suggesting a pivot the paperwork has not caught up with. The domain was registered on 27 September 2024 and first archived on 2 October 2024: this is a young company, and its thin public surface should be read in that light.
Evaluation therefore happens through conversation, not through a trial. Read the articles first, since they are dense enough to tell you quickly whether compiled-output analysis addresses your problem. If it does, write to hello@gestell.ai, ask for the full analysis reports, and use that exchange to settle what the site leaves open: pricing, data hosting, retention, subprocessors, and whether customer artifacts are ever used to train models. For a team maintaining a GPU inference engine, or an enterprise mapping an undocumented mainframe estate, that conversation is likely worth the time.
- Choosing a selection results in a full page refresh.
- Opens in a new window.