Luminal growth playbook

Pick one headline number that the chart on the page supports. Or, if 3.2x is real, show the workload where it was measured: model, hardware, precision, batch size, context length and engine versions. Put a one-line footnote under every stat (3.2x, 2-3x, <10ms p99, 0% overhead) that links to a report page. Each report gives the date, the configuration, the vLLM and TensorRT-LLM versions and settings, and a command to reproduce the run. Add the GPT-OSS 120B run to /report so the homepage chart has a source. The new Head of Growth then gets a first sales asset that says: here is the run, rerun it on your hardware.

Engineers trust a smaller number they can reproduce more than a bigger one they can't. Sourced, consistent claims turn the homepage into something people forward rather than argue with, and next to a reproduce button, a rival's unsourced "up to 4x" looks weaker.

[ ] One headline number that matches the chart
[ ] Model, hardware, precision, batch size, context length
[ ] Engine versions and settings for every baseline
[ ] Date of the run
[ ] Reproduce command and container image
[ ] A link from every homepage stat to its report
[ ] An owner for the quarterly rerun

• The reproducible number is smaller than 3.2x

  A smaller number that holds up beats a bigger one a prospect disproves on a call; lead with cost per token where the gap is larger.

• Baselines look badly tuned

  Publish the baseline settings and invite engine maintainers to suggest better ones.

• Your websiteHomepage (checked twice on November 9, 2026): a headline stat "3.2x" labelled "vs vLLM"; body copy "Models compiled by Luminal consistently outperform existing inference engines by 2-3x on standard benchmarks"; a chart "Tokens/sec, GPT-OSS 120B, 8xH100 SXM" showing Luminal 36k, TensorRT-LLM 28k, vLLM 26k and PyTorch 3k (about 1.4x vLLM and 1.3x TensorRT-LLM); "<10ms" p99 latency and "0%" overhead. None of these has a footnote, date or methodology link, and no customer is named.Source
• Your websiteBenchmarks page (/report): five reports (MiniMax-M2.5, DeepSeek-R1-0528 and Kimi-K2.5 on 8xH200; Qwen3.5-35B-A3B-FP8 and Qwen3-Coder-30B on one H100), each headlined by peak per-user decode speed (93.3 to 172.5 tok/s); no dates, methodology or comparison with vLLM, SGLang or TensorRT-LLM, and no report for the GPT-OSS 120B chart on the homepage.Source
• Competitor researchRivals' own homepages: Fireworks claims "up to 4x the throughput of providers wrapping vLLM", says models are tested against each lab's reference ("KLD <0.007"), names customers such as Notion, Shopify and Cursor, and lists DeepSeek V4.1 Flash at $0.30 input and $1.20 output per million tokens; Together AI claims "up to 3.7x higher throughput" for its Together Compiler with case studies from Cursor, Decagon and Vercept; Baseten sells "The platform for inference, training, and execution".See the research Source
• HiringHead of Growth (Sales, San Francisco, posted November 7, 2026): "Be Luminal's first sales hire and take over our founder-led sales motion"; about 50% pre-sales (pipeline, discovery, closing) and 50% post-sales (onboarding, retention, expansion); 3+ years selling technical infrastructure.Source

Add a strip near the top of the homepage with each live model, its input, cached-input and output prices, measured tokens per second, and a "Get an API key" button. Next to it, add a dated line comparing the price with the rival list price, with the source linked. Do the same for the image model. When the coming text model goes live, add a row instead of rewriting the page.

A developer looking at inference providers can decide in seconds when the price is visible. A lower price is the easiest claim to check, and it backs up the speed story.

• Rivals cut prices

  Date the comparison and lead with Luminal's own price, not the gap.

• Price wars attract low-value users

  Pair the price with speed and the on-prem path for larger teams.

• Your websitePricing page: pay as you go, no subscription; DeepSeek V4.1 Flash live at $0.1425 per million input tokens, $0.0029 for cached input and $0.57 for output; GLM-5.3-Flash "COMING SOON"; Ideogram 4.5 images from $0.008 to $0.20, "the same as Ideogram's own API"; one prepaid balance and one set of API keys for text and images; $50 minimum top-up, credit does not expire, no free tier; default limits of 60 requests and 1,000,000 tokens a minute, with higher limits by emailing sales; no FAQ.Source
• Your websiteHomepage offer: Luminal Cloud ("Managed serverless inference. Deploy in minutes, scale automatically, and only pay for what you use.") and On-Prem ("Licensed cloud or on-prem deployment", "Dedicated engineering support", "Custom kernel optimization", "Strict SLAs tailored to you", "enterprise-grade security"). The homepage names no hosted model and shows no per-token price.Source
• Competitor researchRivals' own homepages: Fireworks claims "up to 4x the throughput of providers wrapping vLLM", says models are tested against each lab's reference ("KLD <0.007"), names customers such as Notion, Shopify and Cursor, and lists DeepSeek V4.1 Flash at $0.30 input and $1.20 output per million tokens; Together AI claims "up to 3.7x higher throughput" for its Together Compiler with case studies from Cursor, Decagon and Vercept; Baseten sells "The platform for inference, training, and execution".See the research Source

Add a second answer to the calculator: Luminal Cloud cost at current per-token prices, with the cached-input price applied to an optional repeat-prompt share. Show both answers side by side, with a crossover point: below this volume use Cloud, above it talk to us about on-prem. Send the Cloud answer to sign-up and the on-prem answer to "Talk to an Engineer".

Visitors get an honest answer for their size and a next step that fits. The on-prem pitch becomes more credible too, because the calculator clearly isn't only selling hardware.

• The calculator's models differ from the hosted model

  Label the Cloud answer with the hosted model, and suggest on-prem when a specific model is required.

• Cannibalising on-prem deals

  Small teams were unlikely to buy 8 GPUs; Cloud keeps them until they cross over.

• Your websiteCalculator: "Right-size your inference infrastructure"; Chat, RAG, Agent or Tools workloads on Qwen3-Coder-30B, MiniMax-M2.5, Kimi-K2.5 or DeepSeek-R1-0528 (H100/H200); at the default 100,000 requests a day (1,024 input and 512 output tokens) it shows OpenAI (GPT-5) $19.2k, Anthropic (Claude Sonnet) $32.3k and Luminal $14.6k a month for 8 GPUs, and "Save up to 55% with Luminal on-prem". It prices on-prem only, not the pay-per-token Cloud on /pricing; CTAs are "Talk to an Engineer", "Get Started" and "How We Calculate".Source
• Your websitePricing page: pay as you go, no subscription; DeepSeek V4.1 Flash live at $0.1425 per million input tokens, $0.0029 for cached input and $0.57 for output; GLM-5.3-Flash "COMING SOON"; Ideogram 4.5 images from $0.008 to $0.20, "the same as Ideogram's own API"; one prepaid balance and one set of API keys for text and images; $50 minimum top-up, credit does not expire, no free tier; default limits of 60 requests and 1,000,000 tokens a minute, with higher limits by emailing sales; no FAQ.Source

Generate a sitemap that lists the homepage, pricing, calculator, /report and every report page, careers, and the blog. Reference it in robots.txt. Add an llms.txt with a short description, the live models and prices, and links to the reports and docs. Give each report page a title and description built from model, hardware and the headline metric.

Each report targets a long-tail query that engineers type when choosing hardware or a provider. A sitemap gets them indexed quickly.

• The content signals restrict AI use

  Decide deliberately which signals to allow; llms.txt only describes public pages.

• Your websiteCrawl files (checked November 9, 2026): luminal.com/robots.txt holds content-signal rules but no Sitemap line, and /sitemap.xml and /llms.txt both return 404.Source
• Your websiteBenchmarks page (/report): five reports (MiniMax-M2.5, DeepSeek-R1-0528 and Kimi-K2.5 on 8xH200; Qwen3.5-35B-A3B-FP8 and Qwen3-Coder-30B on one H100), each headlined by peak per-user decode speed (93.3 to 172.5 tok/s); no dates, methodology or comparison with vLLM, SGLang or TensorRT-LLM, and no report for the GPT-OSS 120B chart on the homepage.Source

Add a short "Run it hosted" section near the top of the README: the live models, prices, one curl example and a sign-up link. Add a Luminal Cloud section to the docs with a quickstart (key, endpoint, first request), limits and pricing, and link the docs header to luminal.com. Use UTM tags so sign-ups from GitHub and the docs can be counted.

Developers who already trust the compiler are the warmest Cloud prospects. One link turns open-source reach into revenue without new spend.

• The open-source community sees it as an upsell

  Keep it short and factual; the compiler stays fully open.

• GitHubluminal-ai/luminal: about 3.0k stars and 254 forks, Apache-2.0/MIT, "a high-performance general-purpose inference compiler" in Rust; CUDA and Metal backends with ROCm on the roadmap; a PyTorch path via torch.compile(model, backend=luminal_cuda); links to docs.luminalai.com, a Discord server and luminal.com, but no mention of Luminal Cloud or its pricing.Source
• DocsDocs at docs.luminalai.com (a different domain from luminal.com): Quickstart, Why Luminal, GraphTensor API, Modules, Compilers, Developers, Blog and Community for the open-source framework; nothing on Luminal Cloud, API keys or pricing, and no link to luminal.com.Source
• Your websitePricing page: pay as you go, no subscription; DeepSeek V4.1 Flash live at $0.1425 per million input tokens, $0.0029 for cached input and $0.57 for output; GLM-5.3-Flash "COMING SOON"; Ideogram 4.5 images from $0.008 to $0.20, "the same as Ideogram's own API"; one prepaid balance and one set of API keys for text and images; $50 minimum top-up, credit does not expire, no free tier; default limits of 60 requests and 1,000,000 tokens a minute, with higher limits by emailing sales; no FAQ.Source

Add a closing block to every post: what this means for your cost, the model price, "Try it with an API key" and "Estimate your cost" links, tagged by post. Link the blog from the site footer and add posts to the sitemap. Over time, mirror posts at luminal.com/blog so search credit builds on the main domain.

Readers of a deep technical post are prospects at peak interest. A clear next step converts some of them, and on-domain posts strengthen luminal.com for search.

• Technical readers dislike sales copy

  Keep it to one factual line and two links.

• BlogBlog on Substack (blog.luminal.com): six posts from November 2025 to November 2026: the seed announcement, "Compiling Models to Megakernels", "Producing The Perfect Token", "Hardware Arbitrage with Luminal and Positron AI", "Reducing Image Generation Cost with AMD and the Luminal Compiler" and "How Much Does A Token Cost?". The latest is a deep study of speculative decoding for the two text models on the pricing page (B300 GPUs, 18 batch sizes, contexts to 32,768) with an interactive explorer hosted off luminal.com; it ends with Substack subscribe prompts and no link to pricing, the calculator or sign-up.Source
• Your websiteCrawl files (checked November 9, 2026): luminal.com/robots.txt holds content-signal rules but no Sitemap line, and /sitemap.xml and /llms.txt both return 404.Source

Pass the calculator inputs and result (GPUs, monthly cost, savings) into the call-booking form, and from there into the CRM. Write a one-page qualification guide for the new hire based on those fields: volume bands, workload types and the Cloud-or-on-prem crossover. Reply within one business day with a short note that restates their numbers.

Founder-led sales depends on context in the founder's head. Putting it in the form gives the new hire that context on day one and shortens discovery.

• The role is not filled yet

  Founders use the same fields until the hire starts.

• HiringHead of Growth (Sales, San Francisco, posted November 7, 2026): "Be Luminal's first sales hire and take over our founder-led sales motion"; about 50% pre-sales (pipeline, discovery, closing) and 50% post-sales (onboarding, retention, expansion); 3+ years selling technical infrastructure.Source
• Your websiteCalculator: "Right-size your inference infrastructure"; Chat, RAG, Agent or Tools workloads on Qwen3-Coder-30B, MiniMax-M2.5, Kimi-K2.5 or DeepSeek-R1-0528 (H100/H200); at the default 100,000 requests a day (1,024 input and 512 output tokens) it shows OpenAI (GPT-5) $19.2k, Anthropic (Claude Sonnet) $32.3k and Luminal $14.6k a month for 8 GPUs, and "Save up to 55% with Luminal on-prem". It prices on-prem only, not the pay-per-token Cloud on /pricing; CTAs are "Talk to an Engineer", "Get Started" and "How We Calculate".Source

Add a standard header to each report: date, Luminal commit, driver and CUDA versions, precision, and a one-command reproduce script. Add a baseline column from the best open engine on the same hardware with published settings. Fix the inconsistent figures and add a changelog. Publish a new report with each model launch.

Reports with baselines answer the buyer's real question: how much faster than what I run now, on my hardware. Dated reports also show steady progress.

• Baselines narrow the gap

  Report cost per token too, where the compiler's edge shows.

• #1 Make the homepage speed claims agree, and link each one to a dated, reproducible benchmark

• Your websiteQwen3.5-35B-A3B-FP8 report: four scenarios from 1,024 (Chatbot) to 32,768 (Tool Calling Agentic) input tokens; peak 172.5 tok/s per user and 9,774 tok/s system throughput at 256 users; safe queue depth falls from 131 (Chatbot) to 8 (Tool Calling Agentic); no date, software versions, reproduction steps or same-hardware baseline; best time to first token is shown as both 0.189 s and 211 ms.Source
• Your websiteBenchmarks page (/report): five reports (MiniMax-M2.5, DeepSeek-R1-0528 and Kimi-K2.5 on 8xH200; Qwen3.5-35B-A3B-FP8 and Qwen3-Coder-30B on one H100), each headlined by peak per-user decode speed (93.3 to 172.5 tok/s); no dates, methodology or comparison with vLLM, SGLang or TensorRT-LLM, and no report for the GPT-OSS 120B chart on the homepage.Source
• HiringEngineering hiring (four roles, San Francisco, posted November 7, 2026): a Cloud Inference Engineer to "Ship low latency, high throughput model serving on Luminal Cloud" (KV caching, paged attention, schedulers, autoscaling, model performance reviews; vLLM, SGLang or TensorRT-LLM experience); a Senior Compiler Engineer (a founding role) to "Develop backend code generation for multiple targets (NVIDIA, Trainium, AMD, TPU, etc.)" and build search-based kernel discovery in Rust; a Founding Kernel Engineer for agent-driven and search-based kernel writing; and a Compiler Engineer.Source

For each hosted model, publish a quality page: divergence from the reference implementation on a fixed prompt set, scores on two or three standard evals, and greedy-decode match rates, all rerun at each compiler release. Link it from pricing and every report. Make it part of the release checklist so a kernel change can't ship if it moves quality.

Removing the hidden-quality-loss objection lets the price and speed claims land. It also gives the kernel team a guardrail that supports faster iteration.

• A measurable drift shows up

  Fix it before publishing; this is the point of the gate.

• Competitor researchRivals' own homepages: Fireworks claims "up to 4x the throughput of providers wrapping vLLM", says models are tested against each lab's reference ("KLD <0.007"), names customers such as Notion, Shopify and Cursor, and lists DeepSeek V4.1 Flash at $0.30 input and $1.20 output per million tokens; Together AI claims "up to 3.7x higher throughput" for its Together Compiler with case studies from Cursor, Decagon and Vercept; Baseten sells "The platform for inference, training, and execution".See the research Source
• BlogBlog on Substack (blog.luminal.com): six posts from November 2025 to November 2026: the seed announcement, "Compiling Models to Megakernels", "Producing The Perfect Token", "Hardware Arbitrage with Luminal and Positron AI", "Reducing Image Generation Cost with AMD and the Luminal Compiler" and "How Much Does A Token Cost?". The latest is a deep study of speculative decoding for the two text models on the pricing page (B300 GPUs, 18 batch sizes, contexts to 32,768) with an interactive explorer hosted off luminal.com; it ends with Substack subscribe prompts and no link to pricing, the calculator or sign-up.Source
• HiringEngineering hiring (four roles, San Francisco, posted November 7, 2026): a Cloud Inference Engineer to "Ship low latency, high throughput model serving on Luminal Cloud" (KV caching, paged attention, schedulers, autoscaling, model performance reviews; vLLM, SGLang or TensorRT-LLM experience); a Senior Compiler Engineer (a founding role) to "Develop backend code generation for multiple targets (NVIDIA, Trainium, AMD, TPU, etc.)" and build search-based kernel discovery in Rust; a Founding Kernel Engineer for agent-driven and search-based kernel writing; and a Compiler Engineer.Source

Give new verified accounts a few dollars of credit, enough for roughly a million tokens and a few images, with the $50 top-up required only after that. Alternatively, test a $10 minimum. Track which test produces more paying accounts after 30 days, not just more sign-ups.

The product's advantages (price and speed) show up in the first few requests. Removing the payment step before that moment lets more developers see them.

• Free-credit abuse

  Require verification and cap credit per card or account.

• Your websitePricing page: pay as you go, no subscription; DeepSeek V4.1 Flash live at $0.1425 per million input tokens, $0.0029 for cached input and $0.57 for output; GLM-5.3-Flash "COMING SOON"; Ideogram 4.5 images from $0.008 to $0.20, "the same as Ideogram's own API"; one prepaid balance and one set of API keys for text and images; $50 minimum top-up, credit does not expire, no free tier; default limits of 60 requests and 1,000,000 tokens a minute, with higher limits by emailing sales; no FAQ.Source

Publish an "Inference for agents" page with a worked example: a 32k-token agent loop with 90% of the prompt repeated, cost per 1,000 runs at Luminal list prices against the normal input price, and how to structure prompts to get cache hits. Add a cache-share slider to the calculator's Cloud mode. Target the page at agent and coding-tool teams.

Agent builders feel input-token cost most. Showing their real cost, not the headline price, makes Luminal much cheaper exactly where it matters to them.

• Cache economics are unsustainable at scale

  Confirm margin with the inference team before promoting.

• #3 Add a Cloud pay-per-token mode to the calculator, which today prices only an 8-GPU on-prem setup

• Your websitePricing page: pay as you go, no subscription; DeepSeek V4.1 Flash live at $0.1425 per million input tokens, $0.0029 for cached input and $0.57 for output; GLM-5.3-Flash "COMING SOON"; Ideogram 4.5 images from $0.008 to $0.20, "the same as Ideogram's own API"; one prepaid balance and one set of API keys for text and images; $50 minimum top-up, credit does not expire, no free tier; default limits of 60 requests and 1,000,000 tokens a minute, with higher limits by emailing sales; no FAQ.Source
• Your websiteQwen3.5-35B-A3B-FP8 report: four scenarios from 1,024 (Chatbot) to 32,768 (Tool Calling Agentic) input tokens; peak 172.5 tok/s per user and 9,774 tok/s system throughput at 256 users; safe queue depth falls from 131 (Chatbot) to 8 (Tool Calling Agentic); no date, software versions, reproduction steps or same-hardware baseline; best time to first token is shown as both 0.189 s and 211 ms.Source
• Your websiteCalculator: "Right-size your inference infrastructure"; Chat, RAG, Agent or Tools workloads on Qwen3-Coder-30B, MiniMax-M2.5, Kimi-K2.5 or DeepSeek-R1-0528 (H100/H200); at the default 100,000 requests a day (1,024 input and 512 output tokens) it shows OpenAI (GPT-5) $19.2k, Anthropic (Claude Sonnet) $32.3k and Luminal $14.6k a month for 8 GPUs, and "Save up to 55% with Luminal on-prem". It prices on-prem only, not the pay-per-token Cloud on /pricing; CTAs are "Talk to an Engineer", "Get Started" and "How We Calculate".Source

Write a "Text and images, one API" page aimed at apps that write and illustrate: marketing tools, e-commerce listings, presentation and design apps. Show a single code sample that does both, the combined price, and the image tiers. Use it in outbound from the new Head of Growth to teams that already use both kinds of model.

Consolidation is a buying reason of its own. Same price plus less admin is an easy yes, and teams that use two products churn less.

• Image volume has thin margin at maker's prices

  Use images to win accounts and earn margin on text.

• Your websitePricing page: pay as you go, no subscription; DeepSeek V4.1 Flash live at $0.1425 per million input tokens, $0.0029 for cached input and $0.57 for output; GLM-5.3-Flash "COMING SOON"; Ideogram 4.5 images from $0.008 to $0.20, "the same as Ideogram's own API"; one prepaid balance and one set of API keys for text and images; $50 minimum top-up, credit does not expire, no free tier; default limits of 60 requests and 1,000,000 tokens a minute, with higher limits by emailing sales; no FAQ.Source
• BlogBlog on Substack (blog.luminal.com): six posts from November 2025 to November 2026: the seed announcement, "Compiling Models to Megakernels", "Producing The Perfect Token", "Hardware Arbitrage with Luminal and Positron AI", "Reducing Image Generation Cost with AMD and the Luminal Compiler" and "How Much Does A Token Cost?". The latest is a deep study of speculative decoding for the two text models on the pricing page (B300 GPUs, 18 batch sizes, contexts to 32,768) with an interactive explorer hosted off luminal.com; it ends with Substack subscribe prompts and no link to pricing, the calculator or sign-up.Source

Replace "COMING SOON" with a "notify me" form. On launch day, publish the price, a benchmark report in the new standard format, and a short blog post, and email the waitlist. Set a standing goal of hosting notable open models within days of release, with a report each time.

Launch-day interest is the cheapest acquisition there is, and a waitlist measures demand before capacity is committed.

• Capacity runs out at launch

  Use the waitlist volume answers to size capacity and stagger invites.

• #8 Give every benchmark report a date, versions, a reproduce command and a same-hardware baseline

• Your websitePricing page: pay as you go, no subscription; DeepSeek V4.1 Flash live at $0.1425 per million input tokens, $0.0029 for cached input and $0.57 for output; GLM-5.3-Flash "COMING SOON"; Ideogram 4.5 images from $0.008 to $0.20, "the same as Ideogram's own API"; one prepaid balance and one set of API keys for text and images; $50 minimum top-up, credit does not expire, no free tier; default limits of 60 requests and 1,000,000 tokens a minute, with higher limits by emailing sales; no FAQ.Source
• BlogBlog on Substack (blog.luminal.com): six posts from November 2025 to November 2026: the seed announcement, "Compiling Models to Megakernels", "Producing The Perfect Token", "Hardware Arbitrage with Luminal and Positron AI", "Reducing Image Generation Cost with AMD and the Luminal Compiler" and "How Much Does A Token Cost?". The latest is a deep study of speculative decoding for the two text models on the pricing page (B300 GPUs, 18 batch sizes, contexts to 32,768) with an interactive explorer hosted off luminal.com; it ends with Substack subscribe prompts and no link to pricing, the calculator or sign-up.Source
• HiringEngineering hiring (four roles, San Francisco, posted November 7, 2026): a Cloud Inference Engineer to "Ship low latency, high throughput model serving on Luminal Cloud" (KV caching, paged attention, schedulers, autoscaling, model performance reviews; vLLM, SGLang or TensorRT-LLM experience); a Senior Compiler Engineer (a founding role) to "Develop backend code generation for multiple targets (NVIDIA, Trainium, AMD, TPU, etc.)" and build search-based kernel discovery in Rust; a Founding Kernel Engineer for agent-driven and search-based kernel writing; and a Compiler Engineer.Source

Publish usage tiers tied to lifetime top-ups (for example $50, $500 and $5,000), each with higher request and token limits, applied automatically. Show the current tier and the next one in the dashboard. Keep email for custom limits above the top tier, routed to the Head of Growth as an expansion lead.

Self-serve tiers remove a stall at the point of adoption, and the top tier becomes a clean signal for sales.

• Capacity spikes from auto-upgraded accounts

  Tie the top tiers to autoscaling headroom.

• Your websitePricing page: pay as you go, no subscription; DeepSeek V4.1 Flash live at $0.1425 per million input tokens, $0.0029 for cached input and $0.57 for output; GLM-5.3-Flash "COMING SOON"; Ideogram 4.5 images from $0.008 to $0.20, "the same as Ideogram's own API"; one prepaid balance and one set of API keys for text and images; $50 minimum top-up, credit does not expire, no free tier; default limits of 60 requests and 1,000,000 tokens a minute, with higher limits by emailing sales; no FAQ.Source
• HiringHead of Growth (Sales, San Francisco, posted November 7, 2026): "Be Luminal's first sales hire and take over our founder-led sales motion"; about 50% pre-sales (pipeline, discovery, closing) and 50% post-sales (onboarding, retention, expansion); 3+ years selling technical infrastructure.Source

Email each paying account a monthly one-pager: tokens and images used, cache-hit share, spend, the same usage at a named rival's published list price, and one tip to cut cost further (for example, prompt structure for cache hits). Flag accounts whose usage is growing quickly to the Head of Growth for an on-prem or committed-volume conversation.

A visible saving each month justifies staying, and growth signals point sales at expansion before a competitor makes an offer.

• Rival prices change

  Date the comparison and use published list prices only.

• HiringHead of Growth (Sales, San Francisco, posted November 7, 2026): "Be Luminal's first sales hire and take over our founder-led sales motion"; about 50% pre-sales (pipeline, discovery, closing) and 50% post-sales (onboarding, retention, expansion); 3+ years selling technical infrastructure.Source
• Your websitePricing page: pay as you go, no subscription; DeepSeek V4.1 Flash live at $0.1425 per million input tokens, $0.0029 for cached input and $0.57 for output; GLM-5.3-Flash "COMING SOON"; Ideogram 4.5 images from $0.008 to $0.20, "the same as Ideogram's own API"; one prepaid balance and one set of API keys for text and images; $50 minimum top-up, credit does not expire, no free tier; default limits of 60 requests and 1,000,000 tokens a minute, with higher limits by emailing sales; no FAQ.Source
• Competitor researchRivals' own homepages: Fireworks claims "up to 4x the throughput of providers wrapping vLLM", says models are tested against each lab's reference ("KLD <0.007"), names customers such as Notion, Shopify and Cursor, and lists DeepSeek V4.1 Flash at $0.30 input and $1.20 output per million tokens; Together AI claims "up to 3.7x higher throughput" for its Together Compiler with case studies from Cursor, Decagon and Vercept; Baseten sells "The platform for inference, training, and execution".See the research Source

Ask three current users (Cloud or on-prem) for a short story: workload, previous setup, cost per token and latency before and after, and time to switch. Where a name can't be published, describe the company by type and size and keep the numbers. Add a customer strip to the homepage and link the stories from pricing and the calculator.

Peer proof is what turns a benchmark into a decision. Each story also tells the new sales hire which use cases to target.

• Customers won't be named

  Publish by company type with verified numbers.

• Your websiteHomepage (checked twice on November 9, 2026): a headline stat "3.2x" labelled "vs vLLM"; body copy "Models compiled by Luminal consistently outperform existing inference engines by 2-3x on standard benchmarks"; a chart "Tokens/sec, GPT-OSS 120B, 8xH100 SXM" showing Luminal 36k, TensorRT-LLM 28k, vLLM 26k and PyTorch 3k (about 1.4x vLLM and 1.3x TensorRT-LLM); "<10ms" p99 latency and "0%" overhead. None of these has a footnote, date or methodology link, and no customer is named.Source
• Competitor researchRivals' own homepages: Fireworks claims "up to 4x the throughput of providers wrapping vLLM", says models are tested against each lab's reference ("KLD <0.007"), names customers such as Notion, Shopify and Cursor, and lists DeepSeek V4.1 Flash at $0.30 input and $1.20 output per million tokens; Together AI claims "up to 3.7x higher throughput" for its Together Compiler with case studies from Cursor, Decagon and Vercept; Baseten sells "The platform for inference, training, and execution".See the research Source
• HiringHead of Growth (Sales, San Francisco, posted November 7, 2026): "Be Luminal's first sales hire and take over our founder-led sales motion"; about 50% pre-sales (pipeline, discovery, closing) and 50% post-sales (onboarding, retention, expansion); 3+ years selling technical infrastructure.Source

Publish one page per rival: per-token price for the shared models, measured throughput and time to first token from a dated run, quality check, limits, on-prem options and how to switch. Keep it factual and sourced, date every row and invite corrections. Start with Fireworks and Together, where a shared model makes the comparison concrete.

Comparison searches are the highest-intent traffic, and a fair page builds the credibility the speed claims need.

• Outdated or disputed numbers

  Date every row, link sources and fix errors quickly.

• #1 Make the homepage speed claims agree, and link each one to a dated, reproducible benchmark
• #9 Publish a quality check proving compiled models give the same answers as the reference

• Press coverageTechCrunch (November 17, 2025): a $5.3 million seed led by Felicis, with angels including Paul Graham and Guillermo Rauch; Y Combinator Summer 2025; Luminal sells compute like CoreWeave and Lambda but competes on compiler optimization; rivals named include Baseten, Together AI, Tensormesh and Clarifai, plus in-house optimization teams at large labs.Source
• Competitor researchRivals' own homepages: Fireworks claims "up to 4x the throughput of providers wrapping vLLM", says models are tested against each lab's reference ("KLD <0.007"), names customers such as Notion, Shopify and Cursor, and lists DeepSeek V4.1 Flash at $0.30 input and $1.20 output per million tokens; Together AI claims "up to 3.7x higher throughput" for its Together Compiler with case studies from Cursor, Decagon and Vercept; Baseten sells "The platform for inference, training, and execution".See the research Source
• Your websitePricing page: pay as you go, no subscription; DeepSeek V4.1 Flash live at $0.1425 per million input tokens, $0.0029 for cached input and $0.57 for output; GLM-5.3-Flash "COMING SOON"; Ideogram 4.5 images from $0.008 to $0.20, "the same as Ideogram's own API"; one prepaid balance and one set of API keys for text and images; $50 minimum top-up, credit does not expire, no free tier; default limits of 60 requests and 1,000,000 tokens a minute, with higher limits by emailing sales; no FAQ.Source

Write a short on-prem pack: supported hardware and drivers, the deployment architecture (air-gapped or not, what leaves the network), update process, SLA tiers and response times, licensing model, and a security questionnaire answered in advance. Put a summary page on the site and share the full pack after the first call.

Enterprise deals stall on unanswered questions. Answering them in advance shortens security review and makes a small team look ready for production.

• Promising more than a small team can support

  Publish only commitments the team can meet today.

• Your websiteHomepage offer: Luminal Cloud ("Managed serverless inference. Deploy in minutes, scale automatically, and only pay for what you use.") and On-Prem ("Licensed cloud or on-prem deployment", "Dedicated engineering support", "Custom kernel optimization", "Strict SLAs tailored to you", "enterprise-grade security"). The homepage names no hosted model and shows no per-token price.Source
• HiringHead of Growth (Sales, San Francisco, posted November 7, 2026): "Be Luminal's first sales hire and take over our founder-led sales motion"; about 50% pre-sales (pipeline, discovery, closing) and 50% post-sales (onboarding, retention, expansion); 3+ years selling technical infrastructure.Source

Launch a "Built with Luminal" page that features community projects, each with a short write-up and a link to the author. Give contributors and featured builders Cloud credit, plus a referral link that earns more credit when a colleague's account starts paying. Run a monthly Discord showcase and pick the best project for the blog.

Recognition and credit cost little and turn the open-source community into a source of referrals and content.

• Few projects to feature at first

  Start with the team's own demos and examples from the repo.

• #5 Point the 3,000-star GitHub repo and the docs at Luminal Cloud

• GitHubluminal-ai/luminal: about 3.0k stars and 254 forks, Apache-2.0/MIT, "a high-performance general-purpose inference compiler" in Rust; CUDA and Metal backends with ROCm on the roadmap; a PyTorch path via torch.compile(model, backend=luminal_cuda); links to docs.luminalai.com, a Discord server and luminal.com, but no mention of Luminal Cloud or its pricing.Source
• DocsDocs at docs.luminalai.com (a different domain from luminal.com): Quickstart, Why Luminal, GraphTensor API, Modules, Compilers, Developers, Blog and Community for the open-source framework; nothing on Luminal Cloud, API keys or pricing, and no link to luminal.com.Source

Publish the Positron prefill and decode benchmark. Then build a joint page for each hardware partner (Positron, AMD, then Trainium and TPU as backends land): supported models, cost per token on that hardware, and a reproduce script, co-announced with the partner. Ask each partner to list Luminal in its own software ecosystem and refer customers.

Hardware vendors have large marketing budgets and customers who need software. Being the compiler that makes their chips usable puts Luminal into their sales conversations.

• Partners want exclusivity

  Stay hardware-neutral; that is the product's value.

• Benchmarks disappoint on early hardware

  Lead with cost per token and the disaggregated setup, where mixed hardware wins.

• #8 Give every benchmark report a date, versions, a reproduce command and a same-hardware baseline

• Blog"Hardware Arbitrage with Luminal and Positron AI" (May 18, 2026): Positron's Atlas added as a compiler backend; prefill compiled for an H100 and decode for Atlas; switching backends is "a single-line code change"; precise prefill and decode benchmarks are promised in a later post, which the blog archive does not list as of November 9, 2026.Source
• HiringEngineering hiring (four roles, San Francisco, posted November 7, 2026): a Cloud Inference Engineer to "Ship low latency, high throughput model serving on Luminal Cloud" (KV caching, paged attention, schedulers, autoscaling, model performance reviews; vLLM, SGLang or TensorRT-LLM experience); a Senior Compiler Engineer (a founding role) to "Develop backend code generation for multiple targets (NVIDIA, Trainium, AMD, TPU, etc.)" and build search-based kernel discovery in Rust; a Founding Kernel Engineer for agent-driven and search-based kernel writing; and a Compiler Engineer.Source
• BlogBlog on Substack (blog.luminal.com): six posts from November 2025 to November 2026: the seed announcement, "Compiling Models to Megakernels", "Producing The Perfect Token", "Hardware Arbitrage with Luminal and Positron AI", "Reducing Image Generation Cost with AMD and the Luminal Compiler" and "How Much Does A Token Cost?". The latest is a deep study of speculative decoding for the two text models on the pricing page (B300 GPUs, 18 batch sizes, contexts to 32,768) with an interactive explorer hosted off luminal.com; it ends with Substack subscribe prompts and no link to pricing, the calculator or sign-up.Source

Each quarter, compile the same few open models for every supported accelerator. Publish cost per million tokens (at public hourly prices), throughput and latency, with scripts to reproduce. Make the index a page on luminal.com with a downloadable dataset, and announce each edition with a blog post and partner quotes.

An index gets cited by press, analysts and buyers each time it updates. It puts Luminal at the centre of the hardware conversation and shows the product working across chips at the same time.

• Seen as biased towards Luminal's partners

  Publish method and scripts and accept outside submissions.

• #20 Publish the promised Positron benchmark and give each accelerator partner a joint page

• Blog"Hardware Arbitrage with Luminal and Positron AI" (May 18, 2026): Positron's Atlas added as a compiler backend; prefill compiled for an H100 and decode for Atlas; switching backends is "a single-line code change"; precise prefill and decode benchmarks are promised in a later post, which the blog archive does not list as of November 9, 2026.Source
• HiringEngineering hiring (four roles, San Francisco, posted November 7, 2026): a Cloud Inference Engineer to "Ship low latency, high throughput model serving on Luminal Cloud" (KV caching, paged attention, schedulers, autoscaling, model performance reviews; vLLM, SGLang or TensorRT-LLM experience); a Senior Compiler Engineer (a founding role) to "Develop backend code generation for multiple targets (NVIDIA, Trainium, AMD, TPU, etc.)" and build search-based kernel discovery in Rust; a Founding Kernel Engineer for agent-driven and search-based kernel writing; and a Compiler Engineer.Source
• BlogBlog on Substack (blog.luminal.com): six posts from November 2025 to November 2026: the seed announcement, "Compiling Models to Megakernels", "Producing The Perfect Token", "Hardware Arbitrage with Luminal and Positron AI", "Reducing Image Generation Cost with AMD and the Luminal Compiler" and "How Much Does A Token Cost?". The latest is a deep study of speculative decoding for the two text models on the pricing page (B300 GPUs, 18 batch sizes, contexts to 32,768) with an interactive explorer hosted off luminal.com; it ends with Substack subscribe prompts and no link to pricing, the calculator or sign-up.Source
• Press coverageTechCrunch (November 17, 2025): a $5.3 million seed led by Felicis, with angels including Paul Graham and Guillermo Rauch; Y Combinator Summer 2025; Luminal sells compute like CoreWeave and Lambda but competes on compiler optimization; rivals named include Baseten, Together AI, Tensormesh and Clarifai, plus in-house optimization teams at large labs.Source

Add a deploy command to the open-source tooling. It takes a model that already compiles locally, uploads the compiled graph, and returns a serverless endpoint on Luminal Cloud with scale-to-zero and per-token or per-second billing. Start with a private beta for teams with fine-tuned or custom models that the hosted catalogue doesn't cover.

It turns every local user of the compiler into a potential Cloud customer, and serves custom models, which the fixed catalogue can't.

• Custom models are harder to serve efficiently

  Start with architectures the compiler already handles well.

• #5 Point the 3,000-star GitHub repo and the docs at Luminal Cloud

• GitHubluminal-ai/luminal: about 3.0k stars and 254 forks, Apache-2.0/MIT, "a high-performance general-purpose inference compiler" in Rust; CUDA and Metal backends with ROCm on the roadmap; a PyTorch path via torch.compile(model, backend=luminal_cuda); links to docs.luminalai.com, a Discord server and luminal.com, but no mention of Luminal Cloud or its pricing.Source
• Your websiteHomepage offer: Luminal Cloud ("Managed serverless inference. Deploy in minutes, scale automatically, and only pay for what you use.") and On-Prem ("Licensed cloud or on-prem deployment", "Dedicated engineering support", "Custom kernel optimization", "Strict SLAs tailored to you", "enterprise-grade security"). The homepage names no hosted model and shows no per-token price.Source
• Your websitePricing page: pay as you go, no subscription; DeepSeek V4.1 Flash live at $0.1425 per million input tokens, $0.0029 for cached input and $0.57 for output; GLM-5.3-Flash "COMING SOON"; Ideogram 4.5 images from $0.008 to $0.20, "the same as Ideogram's own API"; one prepaid balance and one set of API keys for text and images; $50 minimum top-up, credit does not expire, no free tier; default limits of 60 requests and 1,000,000 tokens a minute, with higher limits by emailing sales; no FAQ.Source

Move the explorer to luminal.com/tools/token-cost. Let visitors pick a model, hardware, batch size and context, and see cost per million tokens with and without speculation, alongside Luminal's own price. Add a "Run this on Luminal" button and an embed code so others can cite it. Add more models as they launch.

A useful tool earns links and repeat visits that a blog post doesn't. On luminal.com, those visits build the main site and end one click from sign-up.

• Too technical for most visitors

  Keep the calculator as the simple entry and link the tool for depth.

• BlogBlog on Substack (blog.luminal.com): six posts from November 2025 to November 2026: the seed announcement, "Compiling Models to Megakernels", "Producing The Perfect Token", "Hardware Arbitrage with Luminal and Positron AI", "Reducing Image Generation Cost with AMD and the Luminal Compiler" and "How Much Does A Token Cost?". The latest is a deep study of speculative decoding for the two text models on the pricing page (B300 GPUs, 18 batch sizes, contexts to 32,768) with an interactive explorer hosted off luminal.com; it ends with Substack subscribe prompts and no link to pricing, the calculator or sign-up.Source
• Your websiteCalculator: "Right-size your inference infrastructure"; Chat, RAG, Agent or Tools workloads on Qwen3-Coder-30B, MiniMax-M2.5, Kimi-K2.5 or DeepSeek-R1-0528 (H100/H200); at the default 100,000 requests a day (1,024 input and 512 output tokens) it shows OpenAI (GPT-5) $19.2k, Anthropic (Claude Sonnet) $32.3k and Luminal $14.6k a month for 8 GPUs, and "Save up to 55% with Luminal on-prem". It prices on-prem only, not the pay-per-token Cloud on /pricing; CTAs are "Talk to an Engineer", "Get Started" and "How We Calculate".Source