ProgrammableMarketing/cumuluslabs.io (opens in a new tab)
PDF

Views

  • Plan
  • Roadmap
  • Research

On this page

  • The brief
  • Start here
  • Idea portfolio
  • Effort vs impact
  • This week · 7
  • This quarter · 11
  • Big bets · 3
  • Wildcards · 1
  • How we can help
Generated Nov 10, 2026 1 view
  1. Programmable
  2. Playbooks
  3. Cumulus Labs
Cumulus Labs logo

growth playbook · Nov 10, 2026

cumuluslabs.io (opens in a new tab)

Cumulus Labs growth playbook

Diagnosis

Cumulus Labs is a two-person YC W26 company with unusually deep engineering behind it: a GH200-native inference engine, a GPU orchestrator, an agent platform, 18 serious technical blog posts and a published customer result. The marketing has not caught up with the engineering.

Its strongest claim, 30 to 50 percent more throughput than vLLM or SGLang, appears on its YC profile and Launch HN but not on its own Ion technical page. Pricing sits on another domain, no blog post ends with a next step, the only customer story is filed under Research, and the terms cover the API but not Paladin or Talos.

This playbook first brings the proof, pricing and calls to action onto cumuluslabs.io, then opens migration paths (Retell imports, OpenRouter, fine-tune hosting), and then bets on a public capacity benchmark and outcome-priced agent pilots.

Growth Score

6/10

Biggest Opportunity

Turn the engineering depth into proof buyers can check: a reproducible head-to-head benchmark, a real case study and visible prices, each linked to a clear next step.

Biggest Risk

Well-funded rivals (Fireworks, Together, Baseten, Modal) already sell the same "own your models" story with free credits, SLAs and compliance badges, and a two-person team can lose deals on trust and availability before performance is ever tested.

01

Start here

The three moves to make first. Each one is written up in full below.

  1. 1

    Put the vLLM and SGLang comparison you already claim on the Ion technical page, measured your own open-loop way

    Freeze two workloads (short chat on a 7B model, video on an 8B vision model) and the vLLM and SGLang versions to test against.

    High impact First result in 2 weeks
  2. 2

    Add a Pricing link to the cumuluslabs.io menu that shows the per-token prices now kept in the IonRouter docs

    Copy the current IonRouter model table into a /pricing page and link it from the menu.

    High impact First result in 2 weeks
  3. 3

    End every blog post with the one product step it argues for

    Add a "Get an API key" box to the OpenAI migration post.

    Medium impact First result in 2 weeks

02

Your idea portfolio

22 ideas, grouped by when to run them. Open any idea for how to run it: the steps, who to approach, copy you can use, and how to measure it.

  • 7 ideas are marked Unexpected: plays you are unlikely to have tried.

    1 of these are things Programmable can build and run for you.

Every idea placed by the effort it takes against the impact it could have. Quick wins sit top left; select a number to open that idea. The filters below narrow the list by status, timing, category and what Programmable can build.

Effort vs impactEach number is an idea. Select one to open it.
Impact →
Quick wins · 3Big bets · 8Fill-ins · 7Reconsider · 4
Effort →

Showing 22 of 22 ideas

This week 7

Quick wins with a clear first action.

  • 01

    Cumulus's YC profile says Ion delivers "30 to 50 percent more throughput than standard vLLM or SGLang", and the Launch HN post claimed 588 tok/s against 298 for a large hosted provider. Yet the Ion technical page, where an engineer goes to check, shows only absolute numbers (7,167 tok/s on one GH200; 588 tok/s on video with p50 latency up from 0.74 s to 1.46 s) and no comparison with any other engine. Cumulus's own measurement post explains why peak-throughput benchmarks mislead and promises a "next capacity test" with open-loop arrivals, but publishes no harness. The headline claim lives off-site, and the proof page has no baseline.

    High impactMedium effort First result in 2 weeks CTO

    First action

    Freeze two workloads (short chat on a 7B model, video on an 8B vision model) and the vLLM and SGLang versions to test against.

    How to measure it

    Ion technical page conversions (demo + API signup)

    Now
    –
    Target
    3x the current monthly count
    By
    90 days

    Steps

    1. 1

      Freeze the setup

      Two workloads, model checkpoints, precision, deadlines and the exact vLLM and SGLang releases and flags.

      CTO 1 day

    2. 2

      Build the harness

      Open-loop arrivals, swept request rates, correctness checks; failures and timeouts count as misses, as the measurement post defines.

      Inference Engineer 3 days

    3. 3

      Run all three engines

      Same GH200, repeated boundary points, warm and cold runs.

      Inference Engineer 3 days

    4. 4

      Write it up honestly

      Capacity at deadline, TTFT, p50 and p95, including where Ion loses.

      CTO 2 days

    5. 5

      Publish data and code

      Table on /products/ion/technical, raw CSV and the harness in a public repo.

      CTO 1 day

    6. 6

      Carry the result to the front

      One-line result in the homepage hero and on the Ion page, linked to the full method.

      CEO 1 day

    7. 7

      Add the two next steps

      "Try it on the API" and "Benchmark your workload" with a demo form asking for model and deadline.

      CEO 1 day

    8. 8

      Share it where engineers are

      A follow-up Show HN and posts on X and LinkedIn, with the harness link.

      CEO 1 day

    First result
    2 weeks
    Cost
    A few GPU-days on existing GH200 capacity plus founder and engineer time
    Owner
    CTO

    Smallest test

    Publish one workload (short chat on a 7B model) against vLLM only, with the CSV, and track clicks to the API and demo.

    Why it works

    Run the capacity test the measurement post promises and publish it on /products/ion/technical: the same model, the same GH200 and the same request mix through Ion, current vLLM and current SGLang (versions and flags documented), with an open-loop arrival sweep and capacity reported at a stated deadline and correctness bar. Show the table, the raw CSV and the harness, and include the cases where Ion loses (p50 latency on video). Then put the one-line result ("x% more on-time requests per GPU than vLLM at a 1-second deadline") in the homepage hero and on the Ion page, with two next steps: "Try it on the API" and "Benchmark your workload" (a demo booking that asks for the model and deadline).

    Infrastructure buyers shortlist on numbers they can check, and a comparison is the first thing they look for. Bringing an existing claim onto the site where it can convert, measured with a method that is already published and stricter than the usual peak-throughput charts, answers the Launch HN critic who doubted Cumulus is on the cost and speed frontier. Showing the losses makes the wins believable.

    Ready to use

    ChecklistHead-to-head benchmark publishing list

    [ ] Model, precision, context length and request mix stated
    [ ] Ion, vLLM and SGLang versions and flags listed
    [ ] Open-loop arrival sweep, not fixed concurrency
    [ ] Capacity at a named deadline and correctness bar
    [ ] TTFT, p50 and p95 for each engine
    [ ] Cases where Ion loses shown
    [ ] Raw CSV and harness linked
    [ ] One-line result on the homepage and Ion page
    [ ] "Try it on the API" and "Benchmark your workload" buttons
    [ ] YC profile wording matched to the measured result

    Risks and how to handle them

    • The margin over current vLLM or SGLang is smaller than the YC profile says

      Publish what you measure and update the profile; an honest, reproducible 20% beats an unverifiable 50%.

    • Rival maintainers dispute the configuration

      Publish flags and invite pull requests to the harness; rerun with their suggested settings.

    Evidence

    • YC profileYC W26, founded 2025, team size 2, San Francisco; describes Cumulus as "The Fastest Multimodal Inference OS" and says Ion delivers "30 to 50 percent more throughput than standard vLLM or SGLang"; the launch post says customers save "50-70%" through usage-based pricing.Source
    • Your websiteIon technical page: 7,167 output tokens/s for Qwen2.5-7B-Instruct on one GH200 at 128 concurrent requests, and 588 tokens/s for Qwen3-VL-8B on real video clips with p50 latency of 1.46 s (up from 0.74 s); no comparison with vLLM, SGLang or any other engine or provider appears on the page.Source
    • Launch HN"Launch HN: IonRouter (YC W26) – High-throughput, low-cost inference" (72 points) claimed 588 tok/s against 298 for a large hosted provider on a multimodal workload, while conceding worse p50 latency. Commenters asked for the models/pricing page to be "linked from the top", for zero data retention, for a clear way to host a fine-tune (the founders cited roughly $8-10 per GPU hour with manual onboarding) and for availability on OpenRouter; one wrote it "seems like this is trailing the pareto frontier in cost and speed".Source
    • Your websiteBlog, "Measuring inference throughput under latency constraints" (May 14, 2026): defines capacity as "the highest offered request rate that delivers a stated share of correct responses", uses open-loop arrivals, criticises the vLLM serving benchmark's default infinite request rate, and says the "next capacity test" will use the method; no harness, repository or script is published.Source
  • 02

    cumuluslabs.io has no pricing link anywhere: the homepage shows no prices, and the Ion page sends visitors to a "Models & pricing" anchor inside the IonRouter docs on another domain. The prices themselves are sharp (language models from $0.025 per million input tokens). On Launch HN the first piece of feedback was that the models and pricing page "should be linked from the top".

    High impactLow effort First result in 2 weeks CEO

    First action

    Copy the current IonRouter model table into a /pricing page and link it from the menu.

    How to measure it

    Homepage-to-API-signup rate

    Now
    –
    Target
    +30%
    By
    60 days

    Steps

    1. 1

      Build the page

      Hosted API table, dedicated and own-hardware blocks.

      CTO 1 day

    2. 2

      Single source

      Render prices from the same file the IonRouter docs use.

      CTO 1 day

    3. 3

      Add ranges

      A starting price or typical range for dedicated GPUs.

      CEO half a day

    4. 4

      Write cost examples

      Three workloads priced end to end.

      CEO half a day

    5. 5

      Link it

      Main menu, footer, Ion page and homepage deployment section.

      CTO 1 hour

    6. 6

      Measure

      Clicks from /pricing to signup and demo.

      CEO weekly

    First result
    2 weeks
    Cost
    $0, founder time
    Owner
    CEO

    Smallest test

    Add a menu link to the existing IonRouter pricing anchor today and count clicks.

    Why it works

    Add /pricing on cumuluslabs.io with three blocks: the hosted API table (pulled from the same source as the IonRouter docs so it never drifts), "Dedicated GPUs" with a starting price or a typical monthly range, and "Your infrastructure" with how licensing works. Put Pricing in the main menu and footer, and add a short cost example per workload (for example, one million video-understanding requests).

    Developers comparing inference APIs decide on price first; hiding it on another domain costs the visitors who never click through. A price range for dedicated work also qualifies demo requests.

    Risks and how to handle them

    • Prices on two sites drift apart

      Generate both from one data file.

    • Dedicated ranges anchor buyers too low

      Show "from" prices with what is included.

    Evidence

    • Your websiteHomepage (read November 10, 2026): headline "Beat the frontier. Own your AI."; three products, Paladin (GPU orchestration), Ion (inference) and Talos (agents); calls to action "Get model API access" (to ionrouter.io/login) and "Book a demo" (a Google Calendar booking page); the only proof is "Founded by alumni from" Georgia Tech, NASA, Palantir, Space Force, Blackstone and UW Madison, beside YC W26 and NVIDIA Inception Program badges; no pricing link, benchmark figures, customer names or testimonials.Source
    • Your websiteIon page: "Inference for text, vision, and audio."; to start, create an IonRouter account, add credits and set ION_API_KEY against api.ionrouter.io; hosting as Model API, Dedicated GPUs or Your infrastructure; prices sit behind a "Models & pricing" link to ionrouter.io/docs, and the page itself shows no prices, benchmarks or customers.Source
    • IonRouter docsModels and pricing (read November 10, 2026): language models from $0.025/$0.10 to $0.40/$1.60 per million input/output tokens, vision models at $0.040/$0.25, plus image, video (per GPU-second) and speech (per minute) prices; 100 requests a minute per key; credits must be added before the first call and no free credits are mentioned. A notice reads "Upgrade in progress. Inference is temporarily unavailable." and "Existing customer balances and API keys are preserved; account migration is still in progress."Source
    • Launch HN"Launch HN: IonRouter (YC W26) – High-throughput, low-cost inference" (72 points) claimed 588 tok/s against 298 for a large hosted provider on a multimodal workload, while conceding worse p50 latency. Commenters asked for the models/pricing page to be "linked from the top", for zero data retention, for a clear way to host a fine-tune (the founders cited roughly $8-10 per GPU hour with manual onboarding) and for availability on OpenRouter; one wrote it "seems like this is trailing the pareto frontier in cost and speed".Source
  • 03

    The blog has 18 substantial posts, on GH200 kernels, OpenAI migration, shadow evaluation and voice agents, yet the posts end with share buttons, "Was this helpful?" and related posts. There is no API link, demo link or product link, and no newsletter or RSS feed. Even the OpenAI migration post, whose reader is ready to switch a base URL, has no signup link.

    Medium impactLow effort First result in 2 weeks CEO

    First action

    Add a "Get an API key" box to the OpenAI migration post.

    How to measure it

    Blog-sourced conversions

    Now
    –
    Target
    20 a month
    By
    60 days

    Steps

    1. 1

      Map posts

      One product and one action for each of the 18 posts.

      CEO 2 hours

    2. 2

      Build the component

      A call-to-action box set per post in front matter.

      CTO half a day

    3. 3

      Add email signup

      New-post emails and an RSS feed.

      CTO half a day

    4. 4

      Tag links

      UTM per post so signups are attributed.

      CTO 1 hour

    5. 5

      Ship to all posts

      End of post, and mid-post on the longest.

      CEO 1 day

    6. 6

      Review

      Which posts convert; write more like them.

      CEO monthly

    First result
    2 weeks
    Cost
    $0
    Owner
    CEO

    Smallest test

    Add a box to the three most-read posts and compare clicks over two weeks.

    Why it works

    Map each post to one product and one action: kernel and benchmark posts to "Try Ion on the API", migration to "Get an API key", evaluation and voice posts to "Open Agent Studio" or "Book a Talos demo", GPU idle and scheduling posts to "Review your fleet". Add the box at the end and once mid-post on long pieces, plus an email signup for new posts.

    Technical readers who finish a deep post are the warmest visitors on the site; a specific next step tied to the post's argument converts far better than a generic "contact us".

    Risks and how to handle them

    • Engineers dislike sales boxes

      Keep it one line and technical: "Run this model on Ion".

    Evidence

    • Your websiteBlog: 18 posts from February 10 to September 17, 2026, tagged Product, Engineering, Research or Company; no newsletter signup or RSS feed; the posts read (IonAttention, CoSprite, measurement, OpenAI migration) end with share buttons, "Was this helpful?" and related posts, with no product, API or demo call to action.Source
    • Your websiteBlog, "What OpenAI SDK compatibility requires" (February 26, 2026): Cumulus Gateway, part of Talos, takes an existing OpenAI client with base_url "https://api.cumuluslabs.io/v1", while the Ion page's quickstart uses api.ionrouter.io and ION_API_KEY; the post does not mention IonRouter or link to a signup.Source
  • 04

    "Get model API access" is the homepage's main self-serve button and the first step on the Ion page. On November 10, 2026 the IonRouter docs it leads to said "Upgrade in progress. Inference is temporarily unavailable" and that account migration was still in progress, while cumuluslabs.io said nothing. A developer who signs up and adds credits during that window gets errors with no explanation from the site that sent them, and there is no status page.

    High impactLow effort First result in 1 week CTO

    First action

    Add the status line to the homepage and Ion page buttons today.

    How to measure it

    Signup-to-first-successful-call rate

    Now
    –
    Target
    80%
    By
    30 days

    Steps

    1. 1

      Add one flag

      A single setting for the hosted API state.

      CTO 1 hour

    2. 2

      Show the line

      Next to each "Get model API access" button and on the Ion page.

      CTO 2 hours

    3. 3

      Waitlist

      Email capture when the API is not live.

      CTO 2 hours

    4. 4

      Status page

      Uptime history, incidents and planned work.

      Founding Engineer, ML Platforms 1 day

    5. 5

      Email waiting users

      When service returns, with a credit for the wait.

      CEO 1 hour

    6. 6

      Track first calls

      Share of new keys making a successful request within a day.

      CTO weekly

    First result
    1 week
    Cost
    $0 to $30 a month for a hosted status page
    Owner
    CTO

    Smallest test

    Add the status line to the homepage button and count waitlist signups for a week.

    Build the monitor. A live page fed by the public registries and re-checked every day. Usually live within days.

    Book a working session

    Why it works

    Add a small status line next to every "Get model API access" button and on the Ion page ("Hosted API: live" or "Upgrading, back by <date>; book a dedicated trial meanwhile"), driven by one flag. Publish a simple status page with uptime history and link it from the footer and docs. During any upgrade, send API visitors to a waitlist that emails them when it is back.

    Developers forgive planned downtime; they do not forgive finding it after paying. A visible status line keeps trust, captures the visitors who would otherwise bounce, and gives sales an honest alternative to offer.

    Risks and how to handle them

    • Showing downtime looks weak

      Pair it with a clear return date and a dedicated-trial alternative.

    Evidence

    • IonRouter docsModels and pricing (read November 10, 2026): language models from $0.025/$0.10 to $0.40/$1.60 per million input/output tokens, vision models at $0.040/$0.25, plus image, video (per GPU-second) and speech (per minute) prices; 100 requests a minute per key; credits must be added before the first call and no free credits are mentioned. A notice reads "Upgrade in progress. Inference is temporarily unavailable." and "Existing customer balances and API keys are preserved; account migration is still in progress."Source
    • Your websiteHomepage (read November 10, 2026): headline "Beat the frontier. Own your AI."; three products, Paladin (GPU orchestration), Ion (inference) and Talos (agents); calls to action "Get model API access" (to ionrouter.io/login) and "Book a demo" (a Google Calendar booking page); the only proof is "Founded by alumni from" Georgia Tech, NASA, Palantir, Space Force, Blackstone and UW Madison, beside YC W26 and NVIDIA Inception Program badges; no pricing link, benchmark figures, customer names or testimonials.Source
    • Your websiteIon page: "Inference for text, vision, and audio."; to start, create an IonRouter account, add credits and set ION_API_KEY against api.ionrouter.io; hosting as Model API, Dedicated GPUs or Your infrastructure; prices sit behind a "Models & pricing" link to ionrouter.io/docs, and the page itself shows no prices, benchmarks or customers.Source
  • 05

    The product pages carry exactly the questions buyers type into search and AI assistants: how Paladin differs from Kubernetes, Slurm or NVIDIA Mission Control; "Can tuned models beat frontier models?"; whether to use the API or book a demo. But when the pages were read as text, the headings appeared without their answers, so crawlers and answer engines may see questions with nothing under them.

    Medium impactLow effort First result in 1 month CTO

    First action

    View the Paladin page source and check whether the FAQ answers are in the HTML.

    How to measure it

    Impressions for comparison queries (Kubernetes, Slurm, frontier models)

    Now
    –
    Target
    +50%
    By
    90 days

    Steps

    1. 1

      Check the HTML

      Are answers rendered server-side on Paladin, Ion and Talos pages?

      CTO 1 hour

    2. 2

      Render answers

      Keep the accordion but ship the text in the HTML.

      CTO half a day

    3. 3

      Rewrite answers

      Two or three direct sentences each, with one number where possible.

      CEO 1 day

    4. 4

      Add structured data

      FAQPage markup on each product page.

      CTO 2 hours

    5. 5

      Link deeper

      Each answer links to the technical page or benchmark.

      CTO 1 hour

    6. 6

      Watch queries

      Search Console impressions for the questions.

      CEO monthly

    First result
    1 month
    Cost
    $0
    Owner
    CTO

    Smallest test

    Render the Paladin FAQ in HTML and check it in a text-only fetch.

    Why it works

    Make sure every FAQ answer is in the server-rendered HTML (not only inside a collapsed client-side panel), add FAQPage structured data, and write each answer as a direct, quotable two or three sentences with a link to the technical page or the benchmark. Start with "How is Paladin different from Kubernetes or Slurm?" and "Can tuned models beat frontier models?", the two with the most search demand.

    Comparison questions are high intent, and answer engines quote pages whose answers are plain text. The questions are already written; only the answers need to be visible.

    Risks and how to handle them

    • Answers about rivals go out of date

      Describe Cumulus's approach rather than rivals' features, and date the page.

    Evidence

    • Your websitePaladin pages: "More work. Same GPUs."; predicts each job's needs, packs jobs into spare GPU capacity and rebalances running work, with quotas, permissions and usage visibility, on the customer's fleet or dedicated Cumulus GPUs; FAQ headings include how Paladin differs from Kubernetes, Slurm or NVIDIA Mission Control (their answers did not appear in the page text); neither page gives utilisation, savings or customer figures; the call to action is a 30-minute technical demo ("Review your fleet").Source
    • Your websiteTalos pages: agents for cloud, phone (bring your own SIP trunk and numbers) and edge; "Retell exports can be imported as drafts"; Agent Studio at app.cumuluslabs.io, an API at api.cumuluslabs.io, TypeScript and Python SDKs and MCP; replay, compare and release before rollout; FAQ headings include "Can tuned models beat frontier models?" (answers did not appear in the page text); no pricing; the docs say workspace access is provided by a Cumulus contact.Source
  • 06

    cumuluslabs.io's robots.txt contains two contradicting groups: a Cloudflare-managed block that disallows about 35 AI crawlers, including GPTBot, ClaudeBot and Google-Extended, from the whole site, and Cumulus's own block that explicitly allows GPTBot, ClaudeBot, PerplexityBot and OAI-SearchBot. For a company whose glossary and blog are written to be found by developers asking AI assistants about inference, whether those assistants can read the site is left to each crawler's tie-break rules.

    Medium impactLow effort First result in 1 month CTO

    First action

    Choose which crawlers to allow and remove the conflicting group in Cloudflare.

    How to measure it

    Sessions referred by AI assistants

    Now
    –
    Target
    2x
    By
    90 days

    Steps

    1. 1

      Decide the policy

      Answer engines, search and training crawlers, one by one.

      CEO 1 hour

    2. 2

      Fix Cloudflare

      Turn off or edit the managed robots block.

      CTO 1 hour

    3. 3

      Write one file

      A single group per crawler, with the sitemaps.

      CTO 1 hour

    4. 4

      Test

      Check each named crawler against the live file.

      CTO 1 hour

    5. 5

      Add llms.txt

      Point assistants at products, pricing and the benchmark.

      CTO 2 hours

    6. 6

      Monitor

      AI referrals in analytics, and recheck robots.txt monthly.

      CEO monthly

    First result
    1 month
    Cost
    $0
    Owner
    CTO

    Smallest test

    Remove the conflict for the answer-engine crawlers only and watch referrals.

    Why it works

    Decide the policy (for most, allow answer-engine and search crawlers, block training-only crawlers if wanted), then either turn off Cloudflare's managed block and keep Cumulus's own rules, or edit the managed settings so the two agree. Test each named crawler against the final file and recheck after any Cloudflare change.

    Developers increasingly ask AI assistants which inference provider to use; being readable, and cited, is cheap distribution for a two-person team.

    Risks and how to handle them

    • Allowing training crawlers conflicts with the content-use statement

      Allow search and answer agents, keep training-only agents blocked.

    Evidence

    • Your websiterobots.txt: a Cloudflare-managed block disallows about 35 AI crawlers, including GPTBot, ClaudeBot and Google-Extended, from the whole site, while Cumulus's own block explicitly allows GPTBot, ClaudeBot, PerplexityBot and OAI-SearchBot, so the same crawlers get opposite rules.Source
    • Your websiteSitemaps: /sitemap.xml lists 31 URLs (home, blog, careers, solutions, three product and three technical pages, the glossary and 20 terms), all dated October 9, 2026; /blog/sitemap.xml lists 71: the index, 18 posts and 52 tag pages such as "cheap-gpu", "fast-gpu" and "product%20direction".Source
  • 07

    Cumulus has a published, measured customer result: for CoSprite, "an early Cumulus customer", it built the generation, rendering, selection and fine-tuning pipeline and raised render success from 88% to 94% while halving failed renders. But it is filed as a Research post with no call to action, and the homepage's only proof is the founders' alumni logos, with no customer names or testimonials.

    High impactLow effort First result in 2 weeks CEO

    First action

    Ask CoSprite for a two-sentence quote and approval to feature them.

    How to measure it

    Homepage demo bookings

    Now
    –
    Target
    +20%
    By
    60 days

    Steps

    1. 1

      Get approval

      Quote and logo use from the customer.

      CEO 3 days

    2. 2

      Write the story

      Problem, build, numbers, quote, in 400 words.

      CEO 1 day

    3. 3

      Publish

      A /customers page and the CoSprite story.

      CTO half a day

    4. 4

      Homepage strip

      Customer result under the hero, beside the alumni logos.

      CTO 2 hours

    5. 5

      Link from products

      Talos and Ion pages.

      CTO 1 hour

    6. 6

      Line up the next two

      Ask current production customers for stories.

      CEO 1 month

    First result
    2 weeks
    Cost
    $0
    Owner
    CEO

    Smallest test

    Add one line, "CoSprite: render success 88% to 94%", under the homepage hero.

    Why it works

    Publish a short customer story at /customers/cosprite: the problem, what Cumulus built (Ion inference, fine-tuning, evaluation), the before and after numbers, and a quote from the customer. Put a "Customer results" strip on the homepage under the hero and link it from the Talos page. Ask the next two production customers for the same.

    A named customer with numbers is the strongest proof a young infrastructure company can show, and "companies already run inference and agents on Cumulus in production" is far more convincing with one of them on the page.

    Risks and how to handle them

    • Model-judged results look soft

      Lead with render success, which is measured directly, and note the judge.

    Evidence

    • Your website"Building CoSprite's visual generation pipeline" (March 12, 2026): CoSprite is "an early Cumulus customer" that turns generated code into game previews; Cumulus built the generation, rendering, selection and fine-tuning pipeline, and render success on 50 prompts rose from 88% to 94% (failed renders from 6 to 3); it is filed under Research rather than presented as a case study, and ends with no call to action.Source
    • Your websiteHomepage (read November 10, 2026): headline "Beat the frontier. Own your AI."; three products, Paladin (GPU orchestration), Ion (inference) and Talos (agents); calls to action "Get model API access" (to ionrouter.io/login) and "Book a demo" (a Google Calendar booking page); the only proof is "Founded by alumni from" Georgia Tech, NASA, Palantir, Space Force, Blackstone and UW Madison, beside YC W26 and NVIDIA Inception Program badges; no pricing link, benchmark figures, customer names or testimonials.Source
    • HiringFounding Engineer, ML Platforms Engineer (San Francisco, in person, $150K-$300K, 1-5% equity): build the GPU orchestrator (scheduling, fractional allocation, live workload migration), multi-tenant quotas, isolation, usage metering and a tenant-facing inference gateway, and own fleet observability; Ion is built for GH200, GB200 and GB300; "Companies already run inference and agents on Cumulus in production." The careers page also lists an Inference Engineer for CUDA kernels and model optimisation.Source

This quarter 11

The core plays for the next three months.

  • 08

    Talos already imports Retell exports as drafts without touching the live version, and supports bring-your-own SIP trunks. That is a finished migration path, mentioned only in one line of the technical guide. Retell sells voice agents at $0.07 to $0.31 a minute on models customers cannot tune, and Cumulus's own blog has a post on testing voice agents before migration.

    High impactMedium effort First result in 1 month CEO

    First action

    Write the page from the existing technical guide and the voice-agent post.

    How to measure it

    Retell imports started

    Now
    –
    Target
    15 a month
    By
    1 quarter

    Steps

    1. 1

      Document the importer

      What imports, what does not, known gaps.

      CTO 2 days

    2. 2

      Write the page

      Steps, checks and a cost example per minute.

      CEO 2 days

    3. 3

      Build the replay offer

      Upload 50 saved calls, get quality, latency and cost.

      CTO 1 week

    4. 4

      Check claims

      Use only Retell's published pricing, dated.

      CEO 1 day

    5. 5

      Promote

      Voice-agent developer forums, a blog post and outreach.

      CEO 2 weeks

    6. 6

      Measure

      Imports, replays and conversions to paid.

      CEO monthly

    First result
    1 month
    Cost
    Engineering time plus GPU time for replays
    Owner
    CEO

    Smallest test

    Post the page and offer five free replays to voice-agent teams.

    Why it works

    Build a migration landing page: what imports, what to check (intent, tool calls, state, latency, as the voice-agent post describes), side-by-side replay of saved calls, and a cost comparison per minute once the model is tuned. Offer a free "replay 50 of your calls" check that returns quality, latency and cost against the current setup. Promote it in voice-agent developer communities.

    Switching is the hardest part of selling an agent platform; an importer plus a replayed proof on the buyer's own calls removes most of the risk, and per-minute voice costs make savings easy to show.

    Risks and how to handle them

    • Imported agents behave differently

      Replay before switching and keep the original live until it passes.

    • Compliance-heavy buyers need certifications

      Pair with idea 11's terms and trust page.

    Evidence

    • Your websiteTalos pages: agents for cloud, phone (bring your own SIP trunk and numbers) and edge; "Retell exports can be imported as drafts"; Agent Studio at app.cumuluslabs.io, an API at api.cumuluslabs.io, TypeScript and Python SDKs and MCP; replay, compare and release before rollout; FAQ headings include "Can tuned models beat frontier models?" (answers did not appear in the page text); no pricing; the docs say workspace access is provided by a Cumulus contact.Source
    • Competitor researchRetell: "#1 AI Voice Agent Platform for Automating PHONE Calls"; voice agents at $0.07-$0.31 per minute; new accounts get $10 in free credits and 20 free concurrent calls; HIPAA, SOC2 Type II, GDPR and ISO 27001.Source
    • Your websiteBlog: 18 posts from February 10 to September 17, 2026, tagged Product, Engineering, Research or Company; no newsletter signup or RSS feed; the posts read (IonAttention, CoSprite, measurement, OpenAI migration) end with share buttons, "Was this helpful?" and related posts, with no product, API or demo call to action.Source
  • 09

    A developer reading Cumulus's own material meets two OpenAI-compatible entry points: the Ion page sets ION_API_KEY against api.ionrouter.io, while the OpenAI migration post points the same OpenAI client at api.cumuluslabs.io/v1 through Cumulus Gateway, and Talos uses api.cumuluslabs.io for agents. The migration post never mentions IonRouter. The Founding Engineer role is hired to build "a tenant-facing inference gateway", so the merge is already planned.

    Medium impactMedium effort First result in 1 month Founding Engineer

    First action

    Write the "Start here" page with both endpoints explained side by side.

    How to measure it

    Time to first successful call

    Now
    –
    Target
    under 10 minutes median
    By
    1 quarter

    Steps

    1. 1

      List entry points

      Every base URL, key name and doc page in use.

      Founding Engineer, ML Platforms 1 day

    2. 2

      Write Start here

      One page, two paths, code samples.

      CTO 1 day

    3. 3

      Cross-link docs

      IonRouter docs and docs.cumuluslabs.io point to each other.

      CTO 2 hours

    4. 4

      Plan the merge

      Gateway serves both with one key type.

      Founding Engineer, ML Platforms 1 month

    5. 5

      Keep old URLs

      Old base URL keeps working; docs redirect.

      Founding Engineer, ML Platforms 1 week

    6. 6

      Measure

      Signup to first successful call.

      CTO weekly

    First result
    1 month
    Cost
    Engineering time
    Owner
    Founding Engineer, ML Platforms

    Smallest test

    Add a one-paragraph note to the migration post explaining both endpoints.

    Why it works

    Publish one "Start here" page on cumuluslabs.io: which endpoint to use for plain model calls, which for routed or agent traffic, one key type, and a three-line code sample for each. Until the gateway merges, add a note on both docs sites linking to the other. Once merged, keep the old base URL working and redirect the docs.

    Every extra decision before the first successful call loses developers; one clear path raises the share of new keys that make a request.

    Risks and how to handle them

    • Breaking existing integrations

      Never retire a base URL without months of notice.

    Evidence

    • Your websiteBlog, "What OpenAI SDK compatibility requires" (February 26, 2026): Cumulus Gateway, part of Talos, takes an existing OpenAI client with base_url "https://api.cumuluslabs.io/v1", while the Ion page's quickstart uses api.ionrouter.io and ION_API_KEY; the post does not mention IonRouter or link to a signup.Source
    • Your websiteIon page: "Inference for text, vision, and audio."; to start, create an IonRouter account, add credits and set ION_API_KEY against api.ionrouter.io; hosting as Model API, Dedicated GPUs or Your infrastructure; prices sit behind a "Models & pricing" link to ionrouter.io/docs, and the page itself shows no prices, benchmarks or customers.Source
    • HiringFounding Engineer, ML Platforms Engineer (San Francisco, in person, $150K-$300K, 1-5% equity): build the GPU orchestrator (scheduling, fractional allocation, live workload migration), multi-tenant quotas, isolation, usage metering and a tenant-facing inference gateway, and own fleet observability; Ion is built for GH200, GB200 and GB300; "Companies already run inference and agents on Cumulus in production." The careers page also lists an Inference Engineer for CUDA kernels and model optimisation.Source
  • 10

    On Launch HN, interest in hosting fine-tuned models was strong; the founders quoted roughly $8-10 per GPU hour but onboarding was manual, and one commenter could not find how to host a fine-tune. The IonRouter catalogue today lists only stock models, and the dedicated path is a demo or a founders@ email. Fine-tuned models are also exactly what Talos produces.

    High impactMedium effort First result in 1 month CEO

    First action

    Publish the supported model list and a GPU-hour price, with a request form.

    How to measure it

    Paid fine-tune endpoints

    Now
    –
    Target
    10
    By
    1 quarter

    Steps

    1. 1

      Set the offer

      Supported architectures, sizes and price.

      CEO 2 days

    2. 2

      Publish the page

      On cumuluslabs.io and in the IonRouter docs.

      CTO 1 day

    3. 3

      Form-led provisioning

      Provision within one business day.

      Inference Engineer 1 week

    4. 4

      Automate upload

      Weights upload or registry pull, endpoint created.

      Founding Engineer, ML Platforms 1 month

    5. 5

      Meter usage

      GPU-hours or tokens billed to credits.

      Founding Engineer, ML Platforms 2 weeks

    6. 6

      Link from Talos

      Tuned models deploy in one step.

      CTO 2 weeks

    First result
    1 month
    Cost
    Engineering time; GPU capacity reserved for early endpoints
    Owner
    CEO

    Smallest test

    Publish the price and form, and count requests in two weeks.

    Why it works

    Add a "Bring your weights" page: supported architectures and sizes, upload or registry pull, a published per-GPU-hour or per-token price for dedicated fine-tune endpoints, scale-to-zero rules and expected cold start. Start with a form that provisions within a day, then automate. Cross-sell from Talos fine-tuning so a tuned model goes live in one step.

    Teams with a fine-tune have already decided to use open models and are price-sensitive about serving them; a visible price and a short path capture them before they look elsewhere.

    Risks and how to handle them

    • Idle dedicated endpoints cost money

      Scale to zero with Paladin and price cold starts openly.

    Evidence

    • Launch HN"Launch HN: IonRouter (YC W26) – High-throughput, low-cost inference" (72 points) claimed 588 tok/s against 298 for a large hosted provider on a multimodal workload, while conceding worse p50 latency. Commenters asked for the models/pricing page to be "linked from the top", for zero data retention, for a clear way to host a fine-tune (the founders cited roughly $8-10 per GPU hour with manual onboarding) and for availability on OpenRouter; one wrote it "seems like this is trailing the pareto frontier in cost and speed".Source
    • IonRouter docsModels and pricing (read November 10, 2026): language models from $0.025/$0.10 to $0.40/$1.60 per million input/output tokens, vision models at $0.040/$0.25, plus image, video (per GPU-second) and speech (per minute) prices; 100 requests a minute per key; credits must be added before the first call and no free credits are mentioned. A notice reads "Upgrade in progress. Inference is temporarily unavailable." and "Existing customer balances and API keys are preserved; account migration is still in progress."Source
    • Your websiteTalos pages: agents for cloud, phone (bring your own SIP trunk and numbers) and edge; "Retell exports can be imported as drafts"; Agent Studio at app.cumuluslabs.io, an API at api.cumuluslabs.io, TypeScript and Python SDKs and MCP; replay, compare and release before rollout; FAQ headings include "Can tuned models beat frontier models?" (answers did not appear in the page text); no pricing; the docs say workspace access is provided by a Cumulus contact.Source
  • 11

    The only terms on cumuluslabs.io are the IonRouter Terms of Service, covering the API, playground and docs. They let Cumulus retain outputs and request logs "for a limited period", delete within 30 days on request, offer no zero-retention option and no service level. Paladin runs on customers' fleets and Talos handles customer data on phone and chat, yet neither has its own terms. Launch HN commenters called retention a blocker and asked for zero data retention; Modal and Retell display SOC 2 Type II and HIPAA, and Together offers an uptime SLA.

    High impactMedium effort First result in 1 quarter CEO

    First action

    Decide whether zero retention can be a key-level setting, and write the one-paragraph policy.

    How to measure it

    Days from first demo to signed order

    Now
    –
    Target
    -30%
    By
    2 quarters

    Steps

    1. 1

      Build ZDR

      No storage of prompts or outputs for flagged keys.

      Founding Engineer, ML Platforms 2 weeks

    2. 2

      Draft master terms

      Covering dedicated, Paladin and Talos, with counsel.

      CEO 3 weeks

    3. 3

      Publish a DPA

      Subprocessors and data locations.

      CEO 2 weeks

    4. 4

      Fill the trust center

      Controls, policies and audit status.

      CTO 2 weeks

    5. 5

      Link everywhere

      Footer, pricing and demo pages.

      CTO 1 hour

    6. 6

      Track objections

      Log security questions in each deal.

      CEO ongoing

    First result
    1 quarter
    Cost
    Legal review (a few thousand dollars) and engineering time
    Owner
    CEO

    Smallest test

    Add "Zero data retention available on request" to the pricing page and count requests.

    Why it works

    Add a zero-data-retention option (per key or per workspace) for the hosted API, and publish it on the pricing page. Publish master terms and a data processing agreement that cover dedicated deployments, Paladin and Talos, with an availability target for dedicated customers. Make the Vanta trust center at trust.cumuluslabs.io show what is in progress, and link it from the footer and every demo page.

    Retention and contract questions stop enterprise deals before performance is tested; answering them on the site shortens security review and removes the most-cited objection.

    Risks and how to handle them

    • ZDR limits abuse detection

      Offer it to verified business accounts only.

    Evidence

    • Your websiteThe terms at /terms-of-service.pdf are the "IonRouter.io" Terms of Service (effective March 4, 2026), covering the IonRouter API, playground and docs; Cumulus "may retain Generated Content (images, video, audio, and text outputs) and API request logs for a limited period" and deletes content within 30 days of a request; there is no zero-retention option, no SLA ("we do not guarantee uninterrupted or error-free operation") and no terms for Paladin or Talos; a Trust Center runs at trust.cumuluslabs.io.Source
    • Launch HN"Launch HN: IonRouter (YC W26) – High-throughput, low-cost inference" (72 points) claimed 588 tok/s against 298 for a large hosted provider on a multimodal workload, while conceding worse p50 latency. Commenters asked for the models/pricing page to be "linked from the top", for zero data retention, for a clear way to host a fine-tune (the founders cited roughly $8-10 per GPU hour with manual onboarding) and for availability on OpenRouter; one wrote it "seems like this is trailing the pareto frontier in cost and speed".Source
    • Competitor researchModal: "The platform for production AI"; "$30 / month free compute"; SOC 2 Type II compliant and HIPAA ready.Source
    • Competitor researchRivals' own homepages: Fireworks, "Own your specialized intelligence", claims up to 4x the throughput of providers that wrap vLLM and 10,000+ companies; Together, the "AI Native Cloud" for inference, fine-tuning and GPU clusters, with a 99% uptime SLA on provisioned throughput; Baseten, "The platform for inference, training, and execution".Source
  • 12

    Launch HN commenters asked for IonRouter on OpenRouter and for a single endpoint across many models. Aggregators send traffic by price, speed and uptime, which is the ground Ion competes on, and the IonRouter catalogue already lists prices and approximate speeds per model.

    Medium impactMedium effort First result in 1 month CTO

    First action

    Compare Ion's price and speed with current providers for its top ten models and shortlist five.

    How to measure it

    Monthly tokens via aggregator

    Now
    –
    Target
    1 billion
    By
    2 quarters

    Steps

    1. 1

      Shortlist models

      Where Ion wins on price or speed.

      CTO 2 days

    2. 2

      Check requirements

      Uptime, latency, data-policy and API rules.

      CTO 2 days

    3. 3

      Apply

      Provider application with endpoints and prices.

      CEO 1 day

    4. 4

      Harden capacity

      Rate limits, failover and monitoring for burst traffic.

      Founding Engineer, ML Platforms 2 weeks

    5. 5

      Show the stats

      Link public provider stats from the Ion page.

      CTO 1 hour

    6. 6

      Review weekly

      Share of routed traffic and margin per model.

      CEO weekly

    First result
    1 month
    Cost
    Engineering time; revenue share per the aggregator's terms
    Owner
    CTO

    Smallest test

    List one model and watch routed share for two weeks.

    Why it works

    Pick the five models where Ion's price and measured speed are best, apply to list them as a provider, and meet the aggregator's uptime and latency requirements. Use the listing's public stats as live proof on cumuluslabs.io, and capture users who need dedicated capacity or fine-tunes with a link from the provider page.

    It puts Ion in front of developers already comparing providers, with traffic routed automatically when price and speed are better, without a sales team.

    Risks and how to handle them

    • Low margins on commodity models

      List only models where Ion's cost advantage is real.

    • Outages are publicly visible

      List only after the status work in idea 4 and failover are in place.

    Builds on

    Evidence

    • Launch HN"Launch HN: IonRouter (YC W26) – High-throughput, low-cost inference" (72 points) claimed 588 tok/s against 298 for a large hosted provider on a multimodal workload, while conceding worse p50 latency. Commenters asked for the models/pricing page to be "linked from the top", for zero data retention, for a clear way to host a fine-tune (the founders cited roughly $8-10 per GPU hour with manual onboarding) and for availability on OpenRouter; one wrote it "seems like this is trailing the pareto frontier in cost and speed".Source
    • IonRouter docsModels and pricing (read November 10, 2026): language models from $0.025/$0.10 to $0.40/$1.60 per million input/output tokens, vision models at $0.040/$0.25, plus image, video (per GPU-second) and speech (per minute) prices; 100 requests a minute per key; credits must be added before the first call and no free credits are mentioned. A notice reads "Upgrade in progress. Inference is temporarily unavailable." and "Existing customer balances and API keys are preserved; account migration is still in progress."Source
  • 13

    Paladin's promise is "More work. Same GPUs.", but neither Paladin page shows a utilisation, savings or customer figure, and the only next step is a 30-minute demo where the buyer must bring their fleet configuration. Buyers with GPU fleets know their spend but rarely know how much memory sits idle.

    High impactMedium effort First result in 1 month Founding Engineer

    First action

    Write the export script and a one-page report template from a test fleet.

    How to measure it

    Snapshot-to-demo rate

    Now
    –
    Target
    40%
    By
    1 quarter

    Steps

    1. 1

      Define the inputs

      Which metrics, how long, what format.

      Founding Engineer, ML Platforms 2 days

    2. 2

      Write the exporter

      Read-only script for common setups.

      Founding Engineer, ML Platforms 1 week

    3. 3

      Build the report

      Idle memory, pairs that could share, extra work possible.

      Founding Engineer, ML Platforms 1 week

    4. 4

      Landing page

      On the Paladin page, replacing the bare demo button.

      CTO 2 days

    5. 5

      Run the first ten

      From warm contacts and investors' portfolio companies.

      CEO 1 month

    6. 6

      Publish the aggregate

      Anonymised idle-capacity stat, with consent.

      CEO 1 quarter

    First result
    1 month
    Cost
    Engineering time
    Owner
    Founding Engineer, ML Platforms

    Smallest test

    Offer five snapshots by hand to teams already in conversation.

    Why it works

    Offer a read-only snapshot: the buyer exports a week of standard GPU metrics (memory and utilisation per device and job) with a provided script, and gets back a one-page report: idle memory, jobs that could share a GPU, and the estimated extra work Paladin could pack in. The demo then starts from their number. Aggregate anonymised results into a public "how idle are GPU fleets" stat over time.

    A personal number about wasted spend is a stronger reason to book a call than a product tour, and it qualifies leads by fleet size before the founders spend time.

    Risks and how to handle them

    • Buyers will not share metrics

      Run the script locally and send only the summary.

    Evidence

    • Your websitePaladin pages: "More work. Same GPUs."; predicts each job's needs, packs jobs into spare GPU capacity and rebalances running work, with quotas, permissions and usage visibility, on the customer's fleet or dedicated Cumulus GPUs; FAQ headings include how Paladin differs from Kubernetes, Slurm or NVIDIA Mission Control (their answers did not appear in the page text); neither page gives utilisation, savings or customer figures; the call to action is a 30-minute technical demo ("Review your fleet").Source
    • HiringFounding Engineer, ML Platforms Engineer (San Francisco, in person, $150K-$300K, 1-5% equity): build the GPU orchestrator (scheduling, fractional allocation, live workload migration), multi-tenant quotas, isolation, usage metering and a tenant-facing inference gateway, and own fleet observability; Ion is built for GH200, GB200 and GB300; "Companies already run inference and agents on Cumulus in production." The careers page also lists an Inference Engineer for CUDA kernels and model optimisation.Source
  • 14

    The blog sitemap lists 52 tag pages for only 18 posts, with names such as "cheap-gpu", "fast-gpu" and "product%20direction", so thin near-duplicate pages outnumber articles almost three to one. Meanwhile the 20 glossary terms, which target searches like "LLM router" and "KV cache", link only to each other: the LLM Router entry mentions "Cumulus' Router" once but does not link to IonRouter, any product or any call to action.

    Medium impactLow effort First result in 1 month CTO

    First action

    List the tags with fewer than two posts and noindex them.

    How to measure it

    Glossary-to-product clicks

    Now
    –
    Target
    5% of glossary visits
    By
    90 days

    Steps

    1. 1

      Audit tags

      Posts per tag and traffic.

      CTO 2 hours

    2. 2

      Merge

      Keep 8 to 10 topic tags; redirect or noindex the rest.

      CTO half a day

    3. 3

      Fix the sitemap

      Remove noindexed tags.

      CTO 1 hour

    4. 4

      Map terms to products

      Each of 20 terms to a product page and post.

      CEO 2 hours

    5. 5

      Write the paragraphs

      "How Cumulus does this" with links and one call to action.

      CEO 2 days

    6. 6

      Measure

      Glossary clicks to product pages.

      CTO monthly

    First result
    1 month
    Cost
    $0
    Owner
    CTO

    Smallest test

    Add product links to the five glossary terms closest to Ion.

    Why it works

    Keep 8 to 10 topic tags that match products and themes, noindex or redirect the rest, and drop them from the sitemap. On each glossary term add a "How Cumulus does this" paragraph linking to the matching product, technical page and blog post, plus one call to action (API key or demo).

    Search engines weigh a site by its strongest pages; trimming thin tags concentrates authority, and glossary pages that link to products turn definitional searches into product visits.

    Risks and how to handle them

    • Losing a tag page that ranks

      Check traffic before removing; redirect, do not delete.

    Evidence

    • Your websiteSitemaps: /sitemap.xml lists 31 URLs (home, blog, careers, solutions, three product and three technical pages, the glossary and 20 terms), all dated October 9, 2026; /blog/sitemap.xml lists 71: the index, 18 posts and 52 tag pages such as "cheap-gpu", "fast-gpu" and "product%20direction".Source
    • Your websiteGlossary entry "What is LLM Router?": about 300 words, one passing mention of "Cumulus' Router", no link to IonRouter or any product page, and no call to action; its only links go to other glossary terms.Source
  • 15

    The Solutions page says "Built around your workload." but only repeats the three product cards, with no use cases, industries, numbers or customers. The blog already holds the material for workload pages: voice agents tested before migration, a video VLM benchmark, vision-language-action latency for robots, the CoSprite visual pipeline and GPU sharing across five vision models.

    Medium impactMedium effort First result in 1 month CEO

    First action

    Draft the voice-agent page from the existing post and the Talos guide.

    How to measure it

    Solutions-page demos

    Now
    –
    Target
    10 a month
    By
    1 quarter

    Steps

    1. 1

      Pick four workloads

      Where the blog already has proof.

      CEO 1 day

    2. 2

      Write the pages

      Problem, stack, number, cost, call to action.

      CEO 2 weeks

    3. 3

      Technical review

      Check every number against its source post.

      CTO 2 days

    4. 4

      Build and link

      Pages, homepage links and post links.

      CTO 2 days

    5. 5

      Promote

      Share each page with its community.

      CEO 2 weeks

    6. 6

      Measure

      Visits and demos per page.

      CEO monthly

    First result
    1 month
    Cost
    Founder writing time
    Owner
    CEO

    Smallest test

    Publish the voice-agent page first.

    Why it works

    Turn /solutions into four workload pages, each with the problem, the Cumulus stack that fits (which products, hosted or own hardware), one measured number from the blog, a cost example and one call to action. Link each from the homepage and the relevant blog posts.

    Buyers search for their problem, not product names; workload pages match how they search and show the whole stack working together.

    Risks and how to handle them

    • Too many directions for a small team

      Cut any page that gets no demos in a quarter.

    Evidence

    • Your websiteSolutions page: "Built around your workload." with three product cards (Paladin, Ion, Talos) and no use cases, industries, numbers or customers.Source
    • Your websiteBlog: 18 posts from February 10 to September 17, 2026, tagged Product, Engineering, Research or Company; no newsletter signup or RSS feed; the posts read (IonAttention, CoSprite, measurement, OpenAI migration) end with share buttons, "Was this helpful?" and related posts, with no product, API or demo call to action.Source
  • 16

    IonRouter asks developers to add credits on the Billing page before the first request, and its docs mention no free credits. Rivals let developers try first: Modal advertises "$30 / month free compute", and Retell gives new accounts $10 in free credits. At Ion's prices, a dollar of credit buys millions of tokens on the small models.

    Medium impactLow effort First result in 1 month CTO

    First action

    Set a $2 credit on new accounts for one month as a test.

    How to measure it

    Signup-to-paid rate

    Now
    –
    Target
    +25%
    By
    60 days

    Steps

    1. 1

      Set the amount

      Enough for the quickstart and one real test.

      CEO 1 hour

    2. 2

      Build the grant

      Credit on verified signup.

      Founding Engineer, ML Platforms 2 days

    3. 3

      Add limits

      One per verified identity; lower rate limits until paid.

      Founding Engineer, ML Platforms 1 day

    4. 4

      Usage nudges

      Emails at 50% and 80% used.

      CTO 1 day

    5. 5

      Announce it

      On the pricing page and API button.

      CTO 1 hour

    6. 6

      Compare cohorts

      Paid conversion before and after.

      CEO monthly

    First result
    1 month
    Cost
    Credit cost per signup (a few dollars)
    Owner
    CTO

    Smallest test

    Give credits to half of new signups for a month.

    Why it works

    Give new verified accounts $2 to $5 of credit, enough to run the quickstart and a realistic test on the target model, with a nudge to add a card once 80% is used. Track abuse with simple limits (one per verified email domain or GitHub account).

    Removing the card step before the first call lifts trial; low per-token costs make the credit cheap, and developers who see their own workload run are more likely to buy.

    Risks and how to handle them

    • Abuse by throwaway accounts

      Verification, per-identity limits and low limits until a card is added.

    Builds on

    Evidence

    • IonRouter docsModels and pricing (read November 10, 2026): language models from $0.025/$0.10 to $0.40/$1.60 per million input/output tokens, vision models at $0.040/$0.25, plus image, video (per GPU-second) and speech (per minute) prices; 100 requests a minute per key; credits must be added before the first call and no free credits are mentioned. A notice reads "Upgrade in progress. Inference is temporarily unavailable." and "Existing customer balances and API keys are preserved; account migration is still in progress."Source
    • Competitor researchModal: "The platform for production AI"; "$30 / month free compute"; SOC 2 Type II compliant and HIPAA ready.Source
    • Competitor researchRetell: "#1 AI Voice Agent Platform for Automating PHONE Calls"; voice agents at $0.07-$0.31 per minute; new accounts get $10 in free credits and 20 free concurrent calls; HIPAA, SOC2 Type II, GDPR and ISO 27001.Source
  • 17

    Cumulus published 18 posts in seven months, plus day-zero model support and kernel work, but has no newsletter signup or RSS feed, and the last post is from September 17, 2026. API users who signed up once have no regular reason to come back and see new models or lower prices.

    Medium impactLow effort First result in 1 month CEO

    First action

    Add an email signup to the blog and collect API users' consent.

    How to measure it

    Returning API accounts per month

    Now
    –
    Target
    +20%
    By
    2 quarters

    Steps

    1. 1

      Add signup

      Blog, pricing page and account settings.

      CTO half a day

    2. 2

      Template

      Models, speed, one post, one result.

      CEO 2 hours

    3. 3

      Changelog page

      Same content on cumuluslabs.io.

      CTO half a day

    4. 4

      Send monthly

      First week of each month.

      CEO monthly

    5. 5

      Link usage

      Track keys active after each send.

      CTO monthly

    6. 6

      Refine

      Keep what gets clicked.

      CEO quarterly

    First result
    1 month
    Cost
    $0 to $20 a month for an email tool
    Owner
    CEO

    Smallest test

    Send one letter to existing API users and measure returning keys.

    Why it works

    Send one short monthly email to API accounts and blog subscribers: new models and their prices, measured speed changes, one engineering post, and one customer result. Mirror it as a changelog page on cumuluslabs.io. Use it to announce price cuts and new hosting options to people who already have keys.

    Inference buyers switch providers for price and new models; telling existing users first keeps usage, and a regular cadence keeps a small team visible.

    Risks and how to handle them

    • Founders lack time

      Keep it to five bullet points drawn from commit and release notes.

    Builds on

    Evidence

    • Your websiteBlog: 18 posts from February 10 to September 17, 2026, tagged Product, Engineering, Research or Company; no newsletter signup or RSS feed; the posts read (IonAttention, CoSprite, measurement, OpenAI migration) end with share buttons, "Was this helpful?" and related posts, with no product, API or demo call to action.Source
    • IonRouter docsModels and pricing (read November 10, 2026): language models from $0.025/$0.10 to $0.40/$1.60 per million input/output tokens, vision models at $0.040/$0.25, plus image, video (per GPU-second) and speech (per minute) prices; 100 requests a minute per key; credits must be added before the first call and no free credits are mentioned. A notice reads "Upgrade in progress. Inference is temporarily unavailable." and "Existing customer balances and API keys are preserved; account migration is still in progress."Source
  • 18

    The homepage's GitHub link opens an organisation archived on July 2, 2026, with three archived repositories, none of them Cumulus product code or tooling, and 11 followers. For a developer-facing infrastructure company, GitHub is where engineers check whether a team ships; an archived org says the opposite of a team publishing kernel work every few weeks.

    Medium impactLow effort First result in 1 month CTO

    First action

    Hide or remove the archived repos and pin a new examples repo.

    How to measure it

    Stars on the harness repo

    Now
    –
    Target
    500
    By
    2 quarters

    Steps

    1. 1

      Clean the org

      Hide or remove old experiments.

      CTO 1 hour

    2. 2

      Examples repo

      API, Talos SDK and migration samples.

      CTO 2 days

    3. 3

      Harness repo

      The benchmark from idea 1, with a README.

      Inference Engineer 2 days

    4. 4

      Pin and link

      From docs, blog and homepage.

      CTO 1 hour

    5. 5

      Triage issues

      Answer within two days.

      Inference Engineer weekly

    6. 6

      Release notes

      Tag releases alongside the monthly letter.

      CTO monthly

    First result
    1 month
    Cost
    $0
    Owner
    CTO

    Smallest test

    Publish the examples repo and swap the homepage link.

    Why it works

    Create (or unarchive) an org with three live repos: the open-loop benchmark harness from idea 1, quickstart examples for the API and Talos SDKs in TypeScript and Python, and an OpenAI-migration example. Pin them, link them from the docs and blog, and archive or hide old experiments.

    A small, active GitHub with a useful harness earns stars, forks and issues, which are free developer acquisition and proof of momentum.

    Risks and how to handle them

    • Maintenance load

      Keep it to three repos with clear scope.

    Builds on

    Evidence

    • GitHubThe homepage links to the GitHub org cumulus-compute-labs, which was archived on July 2, 2026: three archived repositories, none of them Cumulus product code or tooling, and 11 followers.Source
    • Your websiteBlog, "Measuring inference throughput under latency constraints" (May 14, 2026): defines capacity as "the highest offered request rate that delivers a stated share of correct responses", uses open-loop arrivals, criticises the vLLM serving benchmark's default infinite request rate, and says the "next capacity test" will use the method; no harness, repository or script is published.Source

Big bets 3

Larger plays that can change the trajectory.

  • 19

    Cumulus has a sharper definition of inference performance than the market uses: capacity as the highest request rate that still returns correct answers within the deadline, with open-loop arrivals, versus the peak-throughput numbers most vendors quote. Fireworks advertises "up to 4x" throughput over vLLM-based providers; buyers have no neutral way to compare. Cumulus is well placed to own the measurement.

    High impactHigh effort First result in 1 quarter CTO

    First action

    Write the public method page and pick three workloads and deadlines.

    How to measure it

    Referring domains to the benchmark

    Now
    –
    Target
    100
    By
    3 quarters

    Steps

    1. 1

      Publish the method

      Workloads, deadlines, correctness checks.

      CTO 2 weeks

    2. 2

      Automate runs

      Monthly, against public APIs at list prices.

      Inference Engineer 1 month

    3. 3

      Legal review

      Provider terms on benchmarking and fair naming.

      CEO 2 weeks

    4. 4

      Launch

      Results page, raw data and code.

      CTO 1 week

    5. 5

      Invite submissions

      Providers can propose configurations.

      CEO ongoing

    6. 6

      Publish monthly

      Results, changes and a short analysis.

      CTO monthly

    First result
    1 quarter
    Cost
    API spend for monthly runs (hundreds of dollars) plus engineering time
    Owner
    CTO

    Smallest test

    Publish one workload across three providers with full data.

    Why it works

    Turn the harness into a public benchmark: a monthly run of the same open-loop workloads (chat, vision, video, speech) against hosted providers' public APIs at their list prices, reported as on-time correct requests per dollar at stated deadlines. Publish the method, code and raw data; invite providers to submit configurations. Include Ion, and say where it loses.

    The company that defines the yardstick gets cited every time it is used. It also produces a steady stream of content and proof for a team with no marketing staff.

    Risks and how to handle them

    • Seen as biased because Cumulus competes

      Open code and data, fixed rules, and outside reviewers.

    • Provider terms restrict benchmarking

      Check each provider's terms before including it.

    Builds on

    Evidence

    • Your websiteBlog, "Measuring inference throughput under latency constraints" (May 14, 2026): defines capacity as "the highest offered request rate that delivers a stated share of correct responses", uses open-loop arrivals, criticises the vLLM serving benchmark's default infinite request rate, and says the "next capacity test" will use the method; no harness, repository or script is published.Source
    • Competitor researchRivals' own homepages: Fireworks, "Own your specialized intelligence", claims up to 4x the throughput of providers that wrap vLLM and 10,000+ companies; Together, the "AI Native Cloud" for inference, fine-tuning and GPU clusters, with a 99% uptime SLA on provisioned throughput; Baseten, "The platform for inference, training, and execution".Source
    • Your websiteIon technical page: 7,167 output tokens/s for Qwen2.5-7B-Instruct on one GH200 at 128 concurrent requests, and 588 tokens/s for Qwen3-VL-8B on real video clips with p50 latency of 1.46 s (up from 0.74 s); no comparison with vLLM, SGLang or any other engine or provider appears on the page.Source
  • 20

    Ion is built for NVIDIA's coherent CPU-GPU memory (GH200, GB200, GB300), and both Ion and Paladin already offer an "on your infrastructure" option. Companies buying Grace systems often run standard serving stacks that do not use the coherent memory, and Cumulus is in NVIDIA's Inception programme. Its strongest technical edge applies wherever those chips are, not only in its own cloud.

    High impactHigh effort First result in 2 quarters CEO

    First action

    Write the licence offer and approach two Grace-system resellers.

    How to measure it

    Licensed GPUs

    Now
    –
    Target
    256
    By
    3 quarters

    Steps

    1. 1

      Define the package

      Ion plus Paladin, installation and support.

      CTO 2 weeks

    2. 2

      Price it

      Per GPU per month, with a pilot price.

      CEO 1 week

    3. 3

      Write the page

      Coherent-memory gains and the benchmark offer.

      CEO 1 week

    4. 4

      Recruit partners

      Resellers, neoclouds and integrators with Grace systems.

      CEO 1 quarter

    5. 5

      Run joint benchmarks

      On each prospect's hardware.

      Inference Engineer per deal

    6. 6

      Package installs

      Repeatable deployment and upgrades.

      Founding Engineer, ML Platforms 1 quarter

    First result
    2 quarters
    Cost
    Founder sales time and engineering for packaging
    Owner
    CEO

    Smallest test

    One paid pilot on a partner's GH200 cluster.

    Why it works

    Package Ion plus Paladin as licensed software for Grace-based fleets, priced per GPU per month. Find buyers through hardware resellers, neoclouds and system integrators that sell GH200 and GB200 systems, offering a joint benchmark on the buyer's hardware before purchase. Publish a "Getting more from your GH200" page with the memory math from the IonAttention post.

    Software licences on customers' hardware scale without Cumulus buying GPUs, and partners who sell the hardware want proof their systems outperform; a joint benchmark gives both sides a reason to sell.

    Risks and how to handle them

    • Support load on many customer environments

      Support a narrow, certified hardware and OS list first.

    Builds on

    Evidence

    • HiringFounding Engineer, ML Platforms Engineer (San Francisco, in person, $150K-$300K, 1-5% equity): build the GPU orchestrator (scheduling, fractional allocation, live workload migration), multi-tenant quotas, isolation, usage metering and a tenant-facing inference gateway, and own fleet observability; Ion is built for GH200, GB200 and GB300; "Companies already run inference and agents on Cumulus in production." The careers page also lists an Inference Engineer for CUDA kernels and model optimisation.Source
    • Your websiteIon page: "Inference for text, vision, and audio."; to start, create an IonRouter account, add credits and set ION_API_KEY against api.ionrouter.io; hosting as Model API, Dedicated GPUs or Your infrastructure; prices sit behind a "Models & pricing" link to ionrouter.io/docs, and the page itself shows no prices, benchmarks or customers.Source
    • Your websitePaladin pages: "More work. Same GPUs."; predicts each job's needs, packs jobs into spare GPU capacity and rebalances running work, with quotas, permissions and usage visibility, on the customer's fleet or dedicated Cumulus GPUs; FAQ headings include how Paladin differs from Kubernetes, Slurm or NVIDIA Mission Control (their answers did not appear in the page text); neither page gives utilisation, savings or customer figures; the call to action is a 30-minute technical demo ("Review your fleet").Source
    • Your websiteHomepage (read November 10, 2026): headline "Beat the frontier. Own your AI."; three products, Paladin (GPU orchestration), Ion (inference) and Talos (agents); calls to action "Get model API access" (to ionrouter.io/login) and "Book a demo" (a Google Calendar booking page); the only proof is "Founded by alumni from" Georgia Tech, NASA, Palantir, Space Force, Blackstone and UW Madison, beside YC W26 and NVIDIA Inception Program badges; no pricing link, benchmark figures, customer names or testimonials.Source
  • 21

    The homepage promise is "Beat the frontier. Own your AI.", Talos asks "Can tuned models beat frontier models?", and Talos can already replay saved conversations and compare a candidate against a baseline before release. The CoSprite result shows the method works on a real customer task. Fireworks now sells the same idea ("Own your specialized intelligence"), so proof per customer is the way to stand out.

    High impactHigh effort First result in 1 quarter CEO

    First action

    Write the pilot terms and success bar, and offer it to two current customers.

    How to measure it

    Pilot-to-contract rate

    Now
    –
    Target
    50%
    By
    2 quarters

    Steps

    1. 1

      Define the pilot

      Length, data needed, success bar and price.

      CEO 1 week

    2. 2

      Standardise the report

      Task success, latency and cost per task.

      CTO 2 weeks

    3. 3

      Write the offer page

      On the Talos page and homepage.

      CEO 1 week

    4. 4

      Run first pilots

      Two to three customers.

      CTO 1 quarter

    5. 5

      Convert

      Contract when the bar is met.

      CEO per pilot

    6. 6

      Publish

      Case studies from passed pilots.

      CEO per pilot

    First result
    1 quarter
    Cost
    GPU time and founder time per pilot, partly covered by the pilot fee
    Owner
    CEO

    Smallest test

    Run one pilot for an existing customer with a written success bar.

    Why it works

    Offer a fixed-length pilot: the customer supplies a sample of real agent traffic; Cumulus tunes a model on Talos, replays the traffic against the customer's current frontier-model setup, and reports task success, latency and cost per task. Price the pilot low, and convert to a contract only if the tuned model meets an agreed bar (for example, equal success at half the cost). Publish each passed pilot (with permission) as a case study.

    Buyers are unsure tuned models can match frontier ones; an outcome-based pilot moves that risk to Cumulus, and the tooling to measure it already exists.

    Risks and how to handle them

    • Tuned model misses the bar

      Choose narrow, repeated tasks first; say no to open-ended ones.

    Builds on

    Evidence

    • Your websiteTalos pages: agents for cloud, phone (bring your own SIP trunk and numbers) and edge; "Retell exports can be imported as drafts"; Agent Studio at app.cumuluslabs.io, an API at api.cumuluslabs.io, TypeScript and Python SDKs and MCP; replay, compare and release before rollout; FAQ headings include "Can tuned models beat frontier models?" (answers did not appear in the page text); no pricing; the docs say workspace access is provided by a Cumulus contact.Source
    • Your website"Building CoSprite's visual generation pipeline" (March 12, 2026): CoSprite is "an early Cumulus customer" that turns generated code into game previews; Cumulus built the generation, rendering, selection and fine-tuning pipeline, and render success on 50 prompts rose from 88% to 94% (failed renders from 6 to 3); it is filed under Research rather than presented as a case study, and ends with no call to action.Source
    • Competitor researchRivals' own homepages: Fireworks, "Own your specialized intelligence", claims up to 4x the throughput of providers that wrap vLLM and 10,000+ companies; Together, the "AI Native Cloud" for inference, fine-tuning and GPU clusters, with a 99% uptime SLA on provisioned throughput; Baseten, "The platform for inference, training, and execution".Source

Wildcards 1

Unexpected experiments, each with a cheap way to test it.

  • 22

    The homepage's only proof is the founders' alumni logos, which include NASA, Space Force and Palantir, and every product already runs on the customer's own infrastructure: Ion on your hardware, Paladin on your fleet, Talos keeping models and customer data where you choose. That is what regulated and public-sector buyers need, but no page speaks to them, and the site already loads fast for real visitors, so the gap is the message, not speed.

    Medium impactMedium effort First result in 1 quarter CEO

    First action

    Draft the page from the existing "your infrastructure" sections of the three product pages.

    How to measure it

    On-premises demos booked

    Now
    –
    Target
    5 a quarter
    By
    2 quarters

    Steps

    1. 1

      Gather the facts

      What each product needs to run inside a customer network.

      CTO 2 days

    2. 2

      Write the page

      Deployment, data control, release controls, pilot.

      CEO 1 week

    3. 3

      Link it

      Homepage proof strip and trust center.

      CTO 1 hour

    4. 4

      List targets

      Labs, universities and integrators in the founders' networks.

      CEO 1 week

    5. 5

      Outreach

      Twenty personal emails offering a pilot.

      CEO 1 month

    6. 6

      Review

      Keep or drop the segment after a quarter.

      CEO quarterly

    First result
    1 quarter
    Cost
    Founder time
    Owner
    CEO

    Smallest test

    Send the draft page to five contacts in research labs and ask if they would take a call.

    Why it works

    Publish one page for teams that cannot send data to a hosted API: the full stack deployed inside their network, the data-control story, the evaluation and release controls in Talos, and how a pilot works. Link it from the homepage proof strip and the trust center. Test it with outreach to research labs, universities and public-sector integrators the founders already know.

    These buyers have budgets, GPU fleets and a hard requirement for on-premises AI; a page addressing them directly, from a team with relevant backgrounds, sets Cumulus apart from hosted-only rivals.

    Risks and how to handle them

    • Long sales cycles for a small team

      Pursue only buyers with an existing GPU fleet and a named project.

    Builds on

    Evidence

    • Your websiteHomepage (read November 10, 2026): headline "Beat the frontier. Own your AI."; three products, Paladin (GPU orchestration), Ion (inference) and Talos (agents); calls to action "Get model API access" (to ionrouter.io/login) and "Book a demo" (a Google Calendar booking page); the only proof is "Founded by alumni from" Georgia Tech, NASA, Palantir, Space Force, Blackstone and UW Madison, beside YC W26 and NVIDIA Inception Program badges; no pricing link, benchmark figures, customer names or testimonials.Source
    • Your websiteTalos pages: agents for cloud, phone (bring your own SIP trunk and numbers) and edge; "Retell exports can be imported as drafts"; Agent Studio at app.cumuluslabs.io, an API at api.cumuluslabs.io, TypeScript and Python SDKs and MCP; replay, compare and release before rollout; FAQ headings include "Can tuned models beat frontier models?" (answers did not appear in the page text); no pricing; the docs say workspace access is provided by a Cumulus contact.Source
    • Your websiteIon page: "Inference for text, vision, and audio."; to start, create an IonRouter account, add credits and set ION_API_KEY against api.ionrouter.io; hosting as Model API, Dedicated GPUs or Your infrastructure; prices sit behind a "Models & pricing" link to ionrouter.io/docs, and the page itself shows no prices, benchmarks or customers.Source
    • PageSpeedChrome UX Report field data for cumuluslabs.io: real visitors' Largest Contentful Paint is 1.6 s at the 75th percentile, INP 84 ms and CLS 0, so the site passes Core Web Vitals; LCP has trended about 14% slower.Source

03

How Programmable can help

Where an idea in this playbook is something Programmable builds or runs, it's listed here with what we'd do. Everything else is yours to run.

Have it built

ProgrammableAgency

1 of the 22 ideas are things Programmable builds and runs for its customers: pages, outreach, prospect lists, booking flows and monitors. Pick one, or send the whole list.

Monitors1A live page fed by the public registries and re-checked every day.

A 30-minute session with the engineer who would build it, not a sales call. You leave with a plan whether or not you hire us.

Book a working session

Usually live within days.

Tools for the rest

ProgrammableDigital

Show it, don't just describe it

These ideas need something to show: a demo, a video, an animation, graphics or a deck. Digital turns screenshots or the site into all of them, in the company's own brand.

Make a demo (opens programmable.digital in a new tab)
More example playbooks and reviews

More growth playbooks

Other companies we have researched, each read from what they are hiring for.

  • Unusual Machines growth strategyDesigns and makes US-built drone components (motors, FPV headsets, cameras, video transmitters, flight controllers and batteries) under…22 growth ideas
  • Atlys growth strategyAn AI-first visa platform that shows the exact visa, fee and delivery date, prepares the paperwork, tracks applications live and…20 growth ideas
  • Routin growth strategyAn AI life tracker that connects tasks, routines, mood, sleep, fitness, health, finance, focus and journaling into one daily Life Score…21 growth ideas
  • Solidgate growth strategyA payment orchestration platform with built-in card acquiring, alternative payment methods, subscription billing, antifraud, dispute…21 growth ideas
  • Verse Medical growth strategyAn AI-enabled DME supplier and ordering platform: clinicians order equipment and supplies online (with details extracted from the medical…20 growth ideas
  • Triade growth strategyA collective of senior independent RevOps, GTM engineering and automation & AI specialists who implement inside B2B teams' HubSpot: CRM…23 growth ideas
  • Lantern growth strategyA specialty care platform (formerly SurgeryPlus) that connects members of self-insured employers, public employers, unions and health…23 growth ideas
  • MMKK AI growth strategyA business technology partner that chooses, sources, sets up and supports business software (Google Workspace, Microsoft 365, Adobe…23 growth ideas

See every playbook and website review

Independent research. Programmable wrote this playbook from public sources, as they stood on November 10, 2026, and cites them throughout. Cumulus Labs did not commission, review or endorse it, and Programmable is not affiliated with Cumulus Labs. Company and product names are trademarks of their owners, used only to identify them. The ideas are suggestions and opinion, not professional, legal or financial advice, and not statements about how Cumulus Labs runs its business.

Something wrong, or do you represent Cumulus Labs? Email [email protected] and we will correct it or take it down.

30/60/90 roadmap

What to focus on in each month, and the ideas that carry it.

  1. Days 1 to 307 ideas

    Put the benchmark comparison, prices, calls to action, API status, FAQ answers and the CoSprite story on cumuluslabs.io, and give crawlers one clear rule.

    Programmable can ship 1 of these in days.

  2. Days 31 to 6011 ideas

    Open migration and self-serve paths: Retell imports, one endpoint quickstart, fine-tune hosting, enterprise terms, OpenRouter, fleet snapshots, workload pages, starter credits, the engineering letter and a live GitHub.

  3. Days 61 to 904 ideas

    Commit to the big bets: a public capacity benchmark, Ion and Paladin on customers' own Grace fleets, outcome-priced agent pilots, and an on-your-hardware path for regulated teams.

The research behind the plan

3 research areas, each summarised in a few lines. Open any of them for the full detail.

Customers

What customers say

Customers value high throughput per GPU and low per-token prices. The most cited problem: idle, expensive GPUs and slow model startup.

  • Drop-in OpenAI-compatible API
  • Not knowing whether a tuned open model can match a frontier model
  • In their words: “tokens per second”

Market

Competitive landscape

5 competitors compared: Fireworks, Together, Baseten, Modal and 1 more.

  • Gap to exploit: Reproducible capacity-at-deadline benchmarks rather than peak throughput
  • Fireworks: Not tied to Grace Hopper memory design
  • Together: Broad catalogue, less workload-specific tuning

Hiring signals

Cumulus is a two-person YC W26 team hiring its first engineers, and both roles are engineering.

  • Founding Engineer, ML Platforms Engineer: Orchestrator, usage metering and a tenant-facing inference gateway. Ideas 9, 13, 16.
  • Inference Engineer, CUDA kernels and model optimization: Kernel and model performance work behind the benchmark. Ideas 1, 19, 20.

From the team that wrote this playbook

All products
  • ProgrammableAgencyWorking AI systems in days, not quarters (opens programmable.agency in a new tab)
  • ProgrammableInfoAn agent that finds, researches and asks firstNew accounts get 100,000 free enrichment credits with code BETA100K (opens programmable.info in a new tab)
  • ProgrammableDigitalScreenshots in, an animated product demo out (opens programmable.digital in a new tab)
  • ProgrammableWatchKnow the day a supplier falls out of compliance (opens programmable.watch in a new tab)
  • ProgrammableTalkGet registrants for your events, and know who’s coming (opens programmable.talk in a new tab)
  • ProgrammableMediaAn AI note taker, interpreter and webinar host for your calls (opens programmable.media in a new tab)
  • ProgrammablePageText what you need, your customer gets a page that works (opens programmable.page in a new tab)
  • ProgrammableCommunityForums with labelled AI members, never empty (opens programmable.community in a new tab)
ProgrammableMarketing
ExamplesLeaderboardsProductsTermsPrivacyCookiesDPALegalContact

© 2026 AI-Powered Growth Intelligence