SF Tensor growth playbook

Add a proof block directly under the H1 of /kernel-optimizer: "#1 overall and in all four tracks of NVIDIA's SOL-ExecBench on B200: 3.03x median speedup over NVIDIA's reference across 235 production kernels, 583 teams", with a link to the write-up and a "Bring us a workload" button beside it. Name and link the benchmark from the homepage card and the labs page, and say which statistic 3.55x is (for example the mean) so every page agrees with the post. End the SOL-ExecBench post with the same "Bring us a workload" button. Tag each booking link with its source page so you can see which pages produce calls.

Kernel and infrastructure buyers check claims before they book a call. A named, externally run benchmark is the strongest proof a young company can show, and the site has no customer logos or case studies to lean on. Putting it on the product page, where the decision is made, takes an afternoon and costs nothing.

#1 on NVIDIA's SOL-ExecBench (B200), as of September 2026
#1 overall and in all four tracks, out of 583 teams and 31,466 submissions.
3.03x median speedup over NVIDIA's optimized reference across 235 production kernels.
Every kernel formally checked against your spec.
[Read how we did it]  [Bring us a workload]

• The benchmark ages as others catch up

  Date the claim ("as of September 2026") and replace it with the next result.

• Mixed figures look careless to technical readers

  Use one headline statistic and say how any other figure was computed.

• Your websiteKernel Optimizer page (title "Automatic Kernel Optimizer | SF Tensor", H1 "Kernels at the speed of light, on any chip."): explains Specify, Search and Prove and says "We run hundreds of millions of tests on each new chip", but never mentions SOL-ExecBench, NVIDIA, B200, FlashAttention, "#1" or any speedup, names no chip vendor, shows no customers or pricing and links to no blog or engineering post. Its calls to action ("Bring us a workload", "Plan a training run", "Talk to an engineer") are all booking links.Source
• Blog"What we learned topping NVIDIA's SOL-ExecBench" (September 25, 2026): SOL-ExecBench is "NVIDIA's official kernel optimization benchmark for the B200" with 235 kernels from production workloads; SF Tensor ranked #1 overall and #1 in every track (average SOL score 0.7992 against 0.7974 for second place), with a 3.03x median speedup per kernel over NVIDIA's optimized reference, among 31,466 submissions from 583 teams. The post also describes an internal suite, KernelWorld, with a held-out slice (KernelWorld-Eval v0.1), and lists weaknesses in the benchmark's grading. It ends with no call to action.Source
• Your websiteAI labs page: "3.55×" "Faster than NVIDIA's baseline on its own kernel benchmark. #1 overall", with the benchmark neither named nor linked; "3× more runs" "From the same compute budget, on average across lab workloads"; 1 to 10,000 GPUs; keep your framework and model code; no customers and no pricing.Source
• Your websiteHomepage (checked November 9, 2026): H1 "Train what doesn't exist yet."; a card headed "#1 on NVIDIA's own benchmark" that names no benchmark and links nowhere; "3x more runs"; up to 10,000 GPUs across NVIDIA, AMD, TPU or Trainium; savings-share pricing ("You only pay us a share of what we save you, which means you pay nothing if your bill doesn't go down") against a self-managed cluster and a managed GPU cloud; an embedded research team with a production model in about six weeks. No customer logos, testimonials or case studies; the only calls to action are two Cal.com booking links ("Plan a training run", "Talk to an engineer"); no self-serve signup, docs or trial.Source

Decide the real list price (a percentage of on-demand, and on which hardware), then make the headline, the body and the table say the same thing. Show Spot as a range rather than one figure, add a short note on what "on-demand" refers to, and add the date. Reuse the same wording anywhere else the tier is described.

Price is the reason to try a new compute class. One consistent, dated figure removes the first objection and makes the post safe for others to quote.

• A lower headline number looks less exciting

  A precise 65% that holds up beats a 70% that a reader can disprove from the same page.

• Blog"Introducing Resumable Compute" (October 29, 2026): "It costs 70% less than on-demand", but the post's own table lists Resumable at "35%" of the on-demand price (65% less) and Spot at "30%", while the text says Spot is "up to 90% cheaper"; billing per second; a multi-node pre-training run moved 13 times across 5 clusters on 3 clouds with a 7-second median gap; "available today in beta on Model Foundry", linking to /model-foundry, with no separate way to request access.Source
• Blog"Solving the Bottleneck to Enterprise AI Sovereignty" (August 4, 2026): argues that compute access, not silicon supply, is the bottleneck, because moving workloads to other chips takes months of manual kernel tuning; claims a "70%" average training cost reduction for enterprise post-training and new-lab pre-training.Source

Add a Resumable Compute section to /model-foundry (what it is, the price, the 13-moves demo, best-fit workloads) with a "Request beta access" button that books a call or opens a short form asking for workload type, GPU count and current provider. Point the launch post's link to that section and tag the source so you can count requests.

Launch posts drive a short burst of intent. If the link lands on the right section with a clear next step, that intent becomes conversations instead of bounces.

• More beta requests than capacity

  Ask about workload size on the form and admit in batches.

• #2 Give Resumable Compute one price: the launch post says 70% less, its own table says 35% of on-demand

• Blog"Introducing Resumable Compute" (October 29, 2026): "It costs 70% less than on-demand", but the post's own table lists Resumable at "35%" of the on-demand price (65% less) and Spot at "30%", while the text says Spot is "up to 90% cheaper"; billing per second; a multi-node pre-training run moved 13 times across 5 clusters on 3 clouds with a 7-second median gap; "available today in beta on Model Foundry", linking to /model-foundry, with no separate way to request access.Source
• Your websiteModel Foundry page: runs launched from Slack, a commit or an agent; the same config on NVIDIA, AMD, TPU or Trainium; lineage across datasets, configs, checkpoints and evals; modules for Datasets, Experiments, Workspaces, Training, Evals, Deploy and Traces. No pricing, no customers, and no mention of Resumable Compute, a beta or a waitlist.Source

Generate sitemap entries for every /engineering/ post and /manifesto with real lastmod dates, and make sure new posts are added automatically. Add article metadata (author, date) to each post, cross-link the engineering and news indexes, and resubmit the sitemap in Search Console.

Technical deep dives earn links and searches for years; making sure they are indexed is the cheapest acquisition work available.

• Pages are already indexed through links

  Still fix it; future posts benefit and lastmod dates help freshness.

• Your websiteSitemap (23 URLs, checked November 9, 2026): the homepage, /enterprise, /labs, /kernel-optimizer, /model-foundry, /news with 8 posts, /engineering and 8 careers URLs. It does not list /manifesto or any of the four /engineering/ posts, and there are no pricing, docs, customer-story or comparison pages.Source
• Your websiteEngineering index: four deep dives, "Formalizing PTX in Lean to Prove GPU Kernel Correctness" (November 2, 2026), "Reverse-Engineering the B200's Tensor Core, One Bit at a Time" (July 11, 2026), "One Model Per Customer Shouldn't Mean One GPU Per Customer" (May 17, 2026) and "We Got Sick of 'Works on My Machine'" (January 15, 2026), at /engineering/formalizing-ptx, /engineering/bitwise-tcgen05, /engineering/multi-tenant-lora and /engineering/infrastructure-tack.Source
• Your websiteManifesto (/manifesto, linked from every footer): "Heterogeneous by default", "Cost is a dimension of performance", "Code should outlive hardware", pointing to the company's own language, Emma; no page on the site describes Emma.Source

Either publish the warpsim post, or point the link at the closest live page (for example the B200 tensor-core post) and remove the reference. Then run a link check across /news and /engineering and add it to the build so it does not happen again.

Small breakages matter more to precise technical readers, and a link checker in the build costs nothing to run.

• The warpsim post is unfinished

  Relink now and restore the link when it ships.

• Blog"Proving Kernels Correct Instead of Testing Them" (August 4, 2026): Z3-based proofs show a FlashAttention-style kernel equivalent to naive attention up to d=128; the post ends with a challenge (prove online softmax in under 100 lines and get an interview) and links to /news/warpsim, which returned 404 Not Found on two checks on November 9, 2026.Source
• Your websiteSitemap (23 URLs, checked November 9, 2026): the homepage, /enterprise, /labs, /kernel-optimizer, /model-foundry, /news with 8 posts, /engineering and 8 careers URLs. It does not list /manifesto or any of the four /engineering/ posts, and there are no pricing, docs, customer-story or comparison pages.Source

Add a two-path footer to every post: "Working on this at your company? Bring us a workload" (booking link tagged with the post) and "Want to work on it? See open roles". Match the offer to the post: a kernel review for the kernel posts, a serving review for the LoRA post, a resumable trial for the Resumable Compute post.

The audience that reads tensor-core bit models includes the engineers who choose training infrastructure. A matched offer turns that attention into conversations without hurting hiring.

• Sales tone puts off technical readers

  Keep it to one plain sentence, framed as an engineering offer.

• Engineering blogB200 tensor-core post (July 11, 2026): a bit-exact model of the tcgen05.mma instruction covering 46 hardware-validated datapaths, released open source at github.com/sf-tensor/tcgen05; it closes by inviting readers who extend the model to apply for a job, with no product call to action.Source
• Blog"Proving Kernels Correct Instead of Testing Them" (August 4, 2026): Z3-based proofs show a FlashAttention-style kernel equivalent to naive attention up to d=128; the post ends with a challenge (prove online softmax in under 100 lines and get an interview) and links to /news/warpsim, which returned 404 Not Found on two checks on November 9, 2026.Source
• Engineering blog"One Model Per Customer Shouldn't Mean One GPU Per Customer" (May 17, 2026, by the CEO): one shared base model with a small LoRA adapter per customer instead of a GPU per customer, citing about 2,000 adapters on a single A100 with S-LoRA; it ends with an invitation to talk.Source
• Blog"What we learned topping NVIDIA's SOL-ExecBench" (September 25, 2026): SOL-ExecBench is "NVIDIA's official kernel optimization benchmark for the B200" with 235 kernels from production workloads; SF Tensor ranked #1 overall and #1 in every track (average SOL score 0.7992 against 0.7974 for second place), with a 3.03x median speedup per kernel over NVIDIA's optimized reference, among 31,466 submissions from 583 teams. The post also describes an internal suite, KernelWorld, with a held-out slice (KernelWorld-Eval v0.1), and lists weaknesses in the benchmark's grading. It ends with no call to action.Source

Add a short header to the README: one line on SF Tensor, links to the write-up, the Kernel Optimizer page and open roles, and a citation block. Pin the repo on the organisation page with a profile README that lists the deep dives. Post the repo, not just the blog, in kernel and compiler communities where people discuss Blackwell numerics.

Developers often find a company through its code before its website. A single link turns each visit to the repo into a possible visit to the site.

• Promotional READMEs annoy developers

  Keep it to one line and links, below the technical summary.

• GitHubOne public repository, tcgen05 (MIT licence, 64 stars, 7 forks, 4 commits); the README has no link to sf-tensor.com or the blog post and no contact, product or hiring prompt.Source
• Engineering blogB200 tensor-core post (July 11, 2026): a bit-exact model of the tcgen05.mma instruction covering 46 hardware-validated datapaths, released open source at github.com/sf-tensor/tcgen05; it closes by inviting readers who extend the model to apply for a job, with no product call to action.Source
• Hacker NewsThe B200 tensor-core post was submitted to Hacker News and drew 3 points with no comments.Source

Add an RSS feed for both indexes and a one-field signup ("Get the next deep dive") at the end of every post and on both index pages. Send one short email per post, and use the list to announce betas such as Resumable Compute first.

An owned list compounds each post's reach and gives every launch a guaranteed first audience, which matters when the team has no marketing hires.

• Low publishing cadence makes the list go cold

  Commit to one email a month, even if it is a short research note.

• Your websiteBlog index: 8 posts from September 2025 to October 29, 2026; no newsletter signup and no RSS feed; LinkedIn is the only social profile in the footer.Source
• Your websiteEngineering index: four deep dives, "Formalizing PTX in Lean to Prove GPU Kernel Correctness" (November 2, 2026), "Reverse-Engineering the B200's Tensor Core, One Bit at a Time" (July 11, 2026), "One Model Per Customer Shouldn't Mean One GPU Per Customer" (May 17, 2026) and "We Got Sick of 'Works on My Machine'" (January 15, 2026), at /engineering/formalizing-ptx, /engineering/bitwise-tcgen05, /engineering/multi-tenant-lora and /engineering/infrastructure-tack.Source
• Hacker NewsThe B200 tensor-core post was submitted to Hacker News and drew 3 points with no comments.Source

Add a dated note at the top: what Tensor Cloud became, with links to Model Foundry and the Resumable Compute post. Keep the original text for the record. Do the same for any other post whose product name has changed.

Old launch posts keep collecting search visits; an update note turns them into a path to the current product instead of a dead end.

• Readers think the product was dropped

  Frame it as "now part of Model Foundry".

• Blog"Tensor Cloud Launch" (October 6, 2025): a private beta of a cloud-agnostic interface that moves jobs off preemptible instances to save "up to 80%"; nothing on the page says whether Tensor Cloud is still a product or has become part of Model Foundry.Source
• Blog"Introducing Resumable Compute" (October 29, 2026): "It costs 70% less than on-demand", but the post's own table lists Resumable at "35%" of the on-demand price (65% less) and Spot at "30%", while the text says Spot is "up to 90% cheaper"; billing per second; a multi-node pre-training run moved 13 times across 5 clusters on 3 clouds with a 7-second median gap; "available today in beta on Model Foundry", linking to /model-foundry, with no separate way to request access.Source

Publish /results: one row per claim with the number, what it measures, the baseline, the workloads, the date and a link to the post or data behind it. Link every headline figure on the site to its row. Update it each quarter, and retire claims that can no longer be backed.

Precise buyers trust numbers they can trace. One page that owns every figure lets the site keep bold claims while removing the doubt that mismatched figures create, and it gives the team a reference for every sales call.

• Some claims cannot be fully backed

  Soften or remove them; fewer strong numbers beat many loose ones.

• Revealing customer workloads

  Describe workloads by type only, with customer approval.

• #1 Put the #1 SOL-ExecBench result on the Kernel Optimizer page, the one page that never mentions it
• #2 Give Resumable Compute one price: the launch post says 70% less, its own table says 35% of on-demand

• Your websiteHomepage (checked November 9, 2026): H1 "Train what doesn't exist yet."; a card headed "#1 on NVIDIA's own benchmark" that names no benchmark and links nowhere; "3x more runs"; up to 10,000 GPUs across NVIDIA, AMD, TPU or Trainium; savings-share pricing ("You only pay us a share of what we save you, which means you pay nothing if your bill doesn't go down") against a self-managed cluster and a managed GPU cloud; an embedded research team with a production model in about six weeks. No customer logos, testimonials or case studies; the only calls to action are two Cal.com booking links ("Plan a training run", "Talk to an engineer"); no self-serve signup, docs or trial.Source
• Your websiteAI labs page: "3.55×" "Faster than NVIDIA's baseline on its own kernel benchmark. #1 overall", with the benchmark neither named nor linked; "3× more runs" "From the same compute budget, on average across lab workloads"; 1 to 10,000 GPUs; keep your framework and model code; no customers and no pricing.Source
• Blog"Solving the Bottleneck to Enterprise AI Sovereignty" (August 4, 2026): argues that compute access, not silicon supply, is the bottleneck, because moving workloads to other chips takes months of manual kernel tuning; claims a "70%" average training cost reduction for enterprise post-training and new-lab pre-training.Source
• Blog"Tensor Cloud Launch" (October 6, 2025): a private beta of a cloud-agnostic interface that moves jobs off preemptible instances to save "up to 80%"; nothing on the page says whether Tensor Cloud is still a product or has become part of Model Foundry.Source
• Your websiteEnterprise page: "Turn proprietary data into proprietary advantage"; financial services, healthcare and life sciences, industrial and public sector; early models by "Week 3" and production in "6 weeks"; specialists at "a fifth of the cost"; customer data "never pooled, shared, or used to train models for anyone else" and training inside the customer's own cloud; a Trust Center link; no named customers, testimonials or case studies.Source

Build a simple estimator on the homepage pricing section: monthly GPU spend, chip type, share of jobs that could run resumable, and current utilisation. It shows a range of expected savings, SF Tensor's share and the net saving, with assumptions listed and linked to /results. The result pre-fills the booking form so the first call starts from numbers.

Savings-share pricing is easiest to sell when the buyer can see the saving before talking to anyone; the estimate also qualifies leads by spend.

• Estimates set expectations that runs miss

  Show a range and the assumptions; savings-share means the customer pays less if savings are lower.

• Publishing the share invites price comparison

  Show the net saving rather than the raw percentage if needed.

• #10 One methodology page behind every headline number: 3x runs, 3.55x, 3.03x, 70%, 80% and "a fifth of the cost"

• Your websiteHomepage (checked November 9, 2026): H1 "Train what doesn't exist yet."; a card headed "#1 on NVIDIA's own benchmark" that names no benchmark and links nowhere; "3x more runs"; up to 10,000 GPUs across NVIDIA, AMD, TPU or Trainium; savings-share pricing ("You only pay us a share of what we save you, which means you pay nothing if your bill doesn't go down") against a self-managed cluster and a managed GPU cloud; an embedded research team with a production model in about six weeks. No customer logos, testimonials or case studies; the only calls to action are two Cal.com booking links ("Plan a training run", "Talk to an engineer"); no self-serve signup, docs or trial.Source
• Competitor researchRivals' own homepages: Modal, "The platform for production AI" with Training and Sandboxes, self-serve signup with $30 a month of free compute, per-second billing and customer logos (DoorDash, Cognition, Ramp); Together AI, "Build what's next on the AI Native Cloud" with GPU clusters and fine-tuning, API signup and named customer results; Lambda, "The Superintelligence Cloud", self-serve GPU instances and SOC 2 Type II.See the research Source
• Blog"Introducing Resumable Compute" (October 29, 2026): "It costs 70% less than on-demand", but the post's own table lists Resumable at "35%" of the on-demand price (65% less) and Spot at "30%", while the text says Spot is "up to 90% cheaper"; billing per second; a multi-node pre-training run moved 13 times across 5 clusters on 3 clouds with a 7-second median gap; "available today in beta on Model Foundry", linking to /model-foundry, with no separate way to request access.Source

Productize the first step: a team sends one PyTorch or JAX operator spec (no data or weights needed), and within 48 hours gets a short report with the fastest kernel found on two target chips, the speedup over their current implementation, and the proof result. Put the offer and a sample report on /kernel-optimizer and /labs, and cap it at a few per week.

A concrete, time-boxed result on the buyer's own operator is far more persuasive than a benchmark, and it starts the sales conversation with evidence.

• Teardowns eat engineering time

  Automate end to end and cap the weekly number.

• Some operators show small gains

  Report honestly and suggest where bigger gains are likely.

• #1 Put the #1 SOL-ExecBench result on the Kernel Optimizer page, the one page that never mentions it

• Your websiteKernel Optimizer page (title "Automatic Kernel Optimizer | SF Tensor", H1 "Kernels at the speed of light, on any chip."): explains Specify, Search and Prove and says "We run hundreds of millions of tests on each new chip", but never mentions SOL-ExecBench, NVIDIA, B200, FlashAttention, "#1" or any speedup, names no chip vendor, shows no customers or pricing and links to no blog or engineering post. Its calls to action ("Bring us a workload", "Plan a training run", "Talk to an engineer") are all booking links.Source
• HiringMember of Technical Staff, GPU Compiler (posted August 29, 2026): "the search found FlashAttention v4-level performance on a B300 in about 25 minutes"; expands "agent- and RL-driven program search" and works "with customer workloads".Source
• HiringMember of Technical Staff, Sandbox Infrastructure (posted August 29, 2026): "a serverless GPU container service across NVIDIA, AMD, TPU and Trainium" that runs compiler measurements, internal post-training and "customer workloads that require isolation".Source
• HiringMember of Mathematical Staff (posted September 27, 2026): will "own the proof engine end-to-end" and work on "formally proving sandboxes safe".Source

Write three short stories, anonymized where needed ("a Series A robotics lab"), each with the starting problem, what SF Tensor changed (kernels, placement, failure recovery), and before-and-after numbers: runs per budget, cost, time to first result. Put one on each of /labs, /enterprise and the homepage, and ask each customer whether they will allow their name later.

Buyers want to see someone like them succeed. Anonymized stories with real numbers give that proof without breaking confidentiality, and they are the material that is missing between the benchmark and the booking.

• Customers decline

  Offer anonymized versions and approval of every line.

• Blog"Reflecting on 3 years of frontier-scale training" (October 19, 2026): the pivot to infrastructure in April 2025; 57 internal tools became Model Foundry; more than 10,000 training runs and experiments and more than 100 unique architectures, on NVIDIA, AMD, TPU, Trainium and 5 other vendors across 23 microarchitectures and 30+ clouds; training for "friends & customers" (none named) in drug discovery, materials science, medical AI, robotics, genomics, image and video generation and voice; thesis: "the labs that win won't have the most compute, they'll have the fastest iteration speed."Source
• Your websiteEnterprise page: "Turn proprietary data into proprietary advantage"; financial services, healthcare and life sciences, industrial and public sector; early models by "Week 3" and production in "6 weeks"; specialists at "a fifth of the cost"; customer data "never pooled, shared, or used to train models for anyone else" and training inside the customer's own cloud; a Trust Center link; no named customers, testimonials or case studies.Source
• Competitor researchRivals' own homepages: Modal, "The platform for production AI" with Training and Sandboxes, self-serve signup with $30 a month of free compute, per-second billing and customer logos (DoorDash, Cognition, Ramp); Together AI, "Build what's next on the AI Native Cloud" with GPU clusters and fine-tuning, API signup and named customer results; Lambda, "The Superintelligence Cloud", self-serve GPU instances and SOC 2 Type II.See the research Source

Create /enterprise/financial-services, /enterprise/healthcare and /enterprise/industrial. Each has one worked example (for instance an underwriting triage model), the data needed, the evaluation used, the week-by-week plan to production, the deployment options and the data boundary. Link them from the enterprise page and use them in outreach.

Industry pages match how enterprise buyers search and give each champion something specific to forward internally.

• Regulated industries need careful claims

  Describe the process and controls, not guaranteed outcomes.

• Your websiteEnterprise page: "Turn proprietary data into proprietary advantage"; financial services, healthcare and life sciences, industrial and public sector; early models by "Week 3" and production in "6 weeks"; specialists at "a fifth of the cost"; customer data "never pooled, shared, or used to train models for anyone else" and training inside the customer's own cloud; a Trust Center link; no named customers, testimonials or case studies.Source
• HiringMember of Technical Staff, Post-Training & Applied Research (posted August 29, 2026): "our enterprise offering promises we'll turn a customer's dataset into a specialist model in days"; designs reward functions with customer domain experts and will "embed with customers as a forward-deployed researcher".Source

Assemble one pack: a diagram of where data, weights and logs live for each deployment option, the ownership and data-use terms in plain language, how sandboxed customer workloads are isolated, answers to the 30 most common security questionnaire items, and the current compliance status from the Trust Center. Offer it on /enterprise and send it after every first enterprise call.

Shortening security review is one of the most effective ways for a small vendor to speed up enterprise deals, and it lets the champion do the internal selling.

• Answers go out of date

  Date the pack and review it each quarter.

• Your websiteEnterprise page: "Turn proprietary data into proprietary advantage"; financial services, healthcare and life sciences, industrial and public sector; early models by "Week 3" and production in "6 weeks"; specialists at "a fifth of the cost"; customer data "never pooled, shared, or used to train models for anyone else" and training inside the customer's own cloud; a Trust Center link; no named customers, testimonials or case studies.Source
• HiringMember of Technical Staff, Sandbox Infrastructure (posted August 29, 2026): "a serverless GPU container service across NVIDIA, AMD, TPU and Trainium" that runs compiler measurements, internal post-training and "customer workloads that require isolation".Source
• Competitor researchRivals' own homepages: Modal, "The platform for production AI" with Training and Sandboxes, self-serve signup with $30 a month of free compute, per-second billing and customer logos (DoorDash, Cognition, Ramp); Together AI, "Build what's next on the AI Native Cloud" with GPU clusters and fine-tuning, API signup and named customer results; Lambda, "The Superintelligence Cloud", self-serve GPU instances and SOC 2 Type II.See the research Source

Generate a one-page report per customer each month inside Model Foundry: runs completed, GPU hours by chip, kernels the search improved and by how much, failures recovered automatically, estimated cost without SF Tensor, the saving and the share billed. Make it easy to forward to finance, and flag workloads that could save more if moved to Resumable Compute or another chip.

A customer who sees the saving every month renews and expands; the report also gives the internal champion numbers to defend the budget.

• Disputes about the baseline cost

  Agree the baseline method at contract start and show it on the report.

• Your websiteHomepage (checked November 9, 2026): H1 "Train what doesn't exist yet."; a card headed "#1 on NVIDIA's own benchmark" that names no benchmark and links nowhere; "3x more runs"; up to 10,000 GPUs across NVIDIA, AMD, TPU or Trainium; savings-share pricing ("You only pay us a share of what we save you, which means you pay nothing if your bill doesn't go down") against a self-managed cluster and a managed GPU cloud; an embedded research team with a production model in about six weeks. No customer logos, testimonials or case studies; the only calls to action are two Cal.com booking links ("Plan a training run", "Talk to an engineer"); no self-serve signup, docs or trial.Source
• HiringMember of Technical Staff, AI-Driven Compilation (posted August 29, 2026): "close the loop with production workloads so that the compiler keeps improving from what customers actually run".Source
• HiringMember of Technical Staff, Product Engineering (posted August 29, 2026, $225k–$275k): "You'll take full ownership of product areas in Model Foundry"; "Foundry is used by both our customers, our internal research team as well as our forward-deployed researchers".Source

Write a one-page brief for investors: why iteration speed decides which new labs win, what SF Tensor changes (more runs per budget, any chip, automatic recovery), the savings-share model (no saving, no fee), and the benchmark result. Send it to the YC network and to funds backing model-training startups, asking for introductions to portfolio teams about to start large runs. Offer portfolio teams a kernel teardown or a free savings estimate.

One investor can introduce many teams at the moment they raise money for compute, and savings-share pricing is easy for an investor to recommend because it carries no downside for the portfolio.

• Investors pass the brief on without context

  Keep it accurate and dated; link to /results.

• Launch YCLaunch YC post: an infrastructure layer for kernel optimization, GPU sourcing and distributed training; the founders describe research turning into "an ugly mix of writing CUDA, debugging driver mismatches and optimizing inter-GPU collective operations"; the ask is introductions to people who train or fine-tune AI models.Source
• Blog"Reflecting on 3 years of frontier-scale training" (October 19, 2026): the pivot to infrastructure in April 2025; 57 internal tools became Model Foundry; more than 10,000 training runs and experiments and more than 100 unique architectures, on NVIDIA, AMD, TPU, Trainium and 5 other vendors across 23 microarchitectures and 30+ clouds; training for "friends & customers" (none named) in drug discovery, materials science, medical AI, robotics, genomics, image and video generation and voice; thesis: "the labs that win won't have the most compute, they'll have the fastest iteration speed."Source

Serve a smaller, preloaded version of the dashboard image to phones, load ad and social tags after the page is interactive, remove any tags no longer used for campaigns, and split unused JavaScript. Recheck PageSpeed and the field data after a month.

Links shared on LinkedIn and messaging apps are often opened on phones; keeping the first view fast protects the impression a technical buyer forms.

• Deferring tags under-counts ad conversions

  Keep conversion events and defer only the libraries.

• PageSpeedGoogle PageSpeed, homepage (November 9, 2026): mobile performance 74 (desktop 99); lab LCP 6.0 s on mobile, where the LCP element is the training-dashboard image; real visitors pass Core Web Vitals (field LCP 1.3 s, CLS 0), though field LCP is 23% slower than in the prior period; about 413 KB of unused JavaScript; Google Tag Manager (275 KB) plus LinkedIn, Reddit and Google ad tags.Source

Publish KernelWorld-Eval as an open benchmark: production-style kernels across several chip vendors, a grader that runs isolated from the candidate, formally checked correctness, and a public leaderboard. Invite other teams and chip vendors to submit. SF Tensor becomes the place where kernel performance is measured across vendors, which suits a company selling portability.

Benchmarks attract the exact audience a kernel company sells to and keep its name in every paper and post that cites the results. Running one across vendors reinforces the "any chip" message.

• Seen as grading its own homework

  Open the grader code and invite outside maintainers.

• Leaks customer workloads

  Use only synthetic or public operators.

• Blog"What we learned topping NVIDIA's SOL-ExecBench" (September 25, 2026): SOL-ExecBench is "NVIDIA's official kernel optimization benchmark for the B200" with 235 kernels from production workloads; SF Tensor ranked #1 overall and #1 in every track (average SOL score 0.7992 against 0.7974 for second place), with a 3.03x median speedup per kernel over NVIDIA's optimized reference, among 31,466 submissions from 583 teams. The post also describes an internal suite, KernelWorld, with a held-out slice (KernelWorld-Eval v0.1), and lists weaknesses in the benchmark's grading. It ends with no call to action.Source
• HiringMember of Mathematical Staff (posted September 27, 2026): will "own the proof engine end-to-end" and work on "formally proving sandboxes safe".Source
• HiringMember of Technical Staff, AI-Driven Compilation (posted August 29, 2026): "close the loop with production workloads so that the compiler keeps improving from what customers actually run".Source

Pick one or two chip vendors or clouds with spare capacity and a reason to win training workloads. Offer a joint program: their sales team brings customers, SF Tensor ports and optimizes the workload with no code changes, and the customer gets the cheaper capacity. Publish one joint benchmark and one joint customer story, and list SF Tensor in the partner's marketplace.

A vendor's sales team reaches far more buyers than a small startup, and the vendor gains a way to remove its biggest objection. Savings-share pricing aligns everyone around the cheaper run.

• Seen as tied to one vendor

  Work with more than one and keep the cross-vendor message.

• Partner leads are low quality

  Agree qualifying criteria up front.

• #12 Turn "Bring us a workload" into a fixed 48-hour kernel teardown with a written report

• Blog"Solving the Bottleneck to Enterprise AI Sovereignty" (August 4, 2026): argues that compute access, not silicon supply, is the bottleneck, because moving workloads to other chips takes months of manual kernel tuning; claims a "70%" average training cost reduction for enterprise post-training and new-lab pre-training.Source
• Blog"Reflecting on 3 years of frontier-scale training" (October 19, 2026): the pivot to infrastructure in April 2025; 57 internal tools became Model Foundry; more than 10,000 training runs and experiments and more than 100 unique architectures, on NVIDIA, AMD, TPU, Trainium and 5 other vendors across 23 microarchitectures and 30+ clouds; training for "friends & customers" (none named) in drug discovery, materials science, medical AI, robotics, genomics, image and video generation and voice; thesis: "the labs that win won't have the most compute, they'll have the fastest iteration speed."Source
• HiringMember of Technical Staff, Sandbox Infrastructure (posted August 29, 2026): "a serverless GPU container service across NVIDIA, AMD, TPU and Trainium" that runs compiler measurements, internal post-training and "customer workloads that require isolation".Source

Launch a limited self-serve tier: sign up, connect a repo or container, run a resumable job on a capped budget with per-second billing, and see the run in Model Foundry. Add short docs for the first job. Teams that grow past the cap are routed to a call for savings-share pricing and the full optimizer.

Self-serve lets researchers prove value to themselves first, widens the top of the funnel beyond people willing to book a call, and feeds qualified, active users to the sales conversation.

• Abuse of free or cheap compute

  Card on file, caps and review of new accounts.

• Distracts from large deals

  Keep self-serve narrow: resumable jobs only.

• #3 Give the Resumable Compute beta a way in: the launch post sends readers to a Model Foundry page that never mentions it
• #8 Start an email list and RSS feed so deep-dive readers come back for the next result

• Your websiteHomepage (checked November 9, 2026): H1 "Train what doesn't exist yet."; a card headed "#1 on NVIDIA's own benchmark" that names no benchmark and links nowhere; "3x more runs"; up to 10,000 GPUs across NVIDIA, AMD, TPU or Trainium; savings-share pricing ("You only pay us a share of what we save you, which means you pay nothing if your bill doesn't go down") against a self-managed cluster and a managed GPU cloud; an embedded research team with a production model in about six weeks. No customer logos, testimonials or case studies; the only calls to action are two Cal.com booking links ("Plan a training run", "Talk to an engineer"); no self-serve signup, docs or trial.Source
• Competitor researchRivals' own homepages: Modal, "The platform for production AI" with Training and Sandboxes, self-serve signup with $30 a month of free compute, per-second billing and customer logos (DoorDash, Cognition, Ramp); Together AI, "Build what's next on the AI Native Cloud" with GPU clusters and fine-tuning, API signup and named customer results; Lambda, "The Superintelligence Cloud", self-serve GPU instances and SOC 2 Type II.See the research Source
• HiringMember of Technical Staff, Sandbox Infrastructure (posted August 29, 2026): "a serverless GPU container service across NVIDIA, AMD, TPU and Trainium" that runs compiler measurements, internal post-training and "customer workloads that require isolation".Source

Collect the challenges on one /challenges page with a public leaderboard, a submission repo and a monthly new problem (a kernel to beat, a proof to write). Ask entrants one optional question: "Does your team train models?" and offer those teams a kernel teardown. Feature winners in the newsletter.

One page serves hiring and acquisition at once and gives engineers a reason to return; it also produces shareable content without a marketing team.

• Mixing recruiting and sales feels pushy

  Keep the sales question optional and separate.

• #8 Start an email list and RSS feed so deep-dive readers come back for the next result

• Blog"Proving Kernels Correct Instead of Testing Them" (August 4, 2026): Z3-based proofs show a FlashAttention-style kernel equivalent to naive attention up to d=128; the post ends with a challenge (prove online softmax in under 100 lines and get an interview) and links to /news/warpsim, which returned 404 Not Found on two checks on November 9, 2026.Source
• Engineering blogB200 tensor-core post (July 11, 2026): a bit-exact model of the tcgen05.mma instruction covering 46 hardware-validated datapaths, released open source at github.com/sf-tensor/tcgen05; it closes by inviting readers who extend the model to apply for a job, with no product call to action.Source
• HiringCareers page: 7 open roles, all on-site in San Francisco and all in engineering or research, none in sales or marketing; the team writes "the compiler, the kernels and the runtime that decide whether a training run costs 1x or 7x".Source

Publish /emma with what the language is, how it relates to the PyTorch and JAX specs the optimizer accepts, one small example (an attention or normalization operator) and its output on two chip vendors, and a waitlist for early access. If Emma is not ready to show, reword the manifesto to describe the goal without naming it.

A language is a strong developer magnet; even a small public preview gives engineers something to discuss and share, and an honest manifesto avoids a dead end.

• Showing an unfinished language invites criticism

  Label it a preview and show one solid example.

• #4 Add the manifesto and the four engineering deep dives to the sitemap

• Your websiteManifesto (/manifesto, linked from every footer): "Heterogeneous by default", "Cost is a dimension of performance", "Code should outlive hardware", pointing to the company's own language, Emma; no page on the site describes Emma.Source
• Your websiteSitemap (23 URLs, checked November 9, 2026): the homepage, /enterprise, /labs, /kernel-optimizer, /model-foundry, /news with 8 posts, /engineering and 8 careers URLs. It does not list /manifesto or any of the four /engineering/ posts, and there are no pricing, docs, customer-story or comparison pages.Source