https://nextlm.ai/assets/ArticlePage-Bv_S-4aO.js

1import{bL as D,u as G,a as V,j as e,g as L,e as C,aE as w,F as v,b3 as S}from"./index-ClCfwAVK.js";import{b as h,e as M}from"./vendor-recharts-d-tUjms2.js";import{a as j,b as N,Y as E}from"./image0_399_2726-XfF2Q7bp.js";import{P as R,a as W,b as P,B as y}from"./PublicLayout-DXsgUEDK.js";import{S as B}from"./share-2-DV7u2r7q.js";import{G as U}from"./graduation-cap-DPISn78b.js";import"./vendor-chart-core-B41Fk490.js";import"./captureVisitorEmail-DqU1xXYh.js";import"./signupHandoff-CFOqdPM6.js";const O="/assets/hero-band-B5kQ9snE.webp",z="/assets/cta-band-Bs56l0oc.webp",X="/assets/card-intent-CapOxpyi.webp",$="/assets/card-playbook-WW5OrXbF.webp",q="/assets/card-signals-Es3VNL4G.webp",F="/assets/card-coldleads--03ilNYw.webp",_="/assets/card-enterprise-CyGLYAzl.webp",K="/assets/card-learning-CJR4xAp6.webp",o={slug:"goliath",category:"Perspective",title:"Your pipeline has a Goliath problem.",description:"The systems built to find every possible account have made it harder to see the few people ready to move. Here is how revenue teams can get small, precise, and early.",author:"NextLM Editorial",authorRole:"Research and market intelligence",publishedAt:"2026-08-26",publishedLabel:"August 26, 2026",readTime:"7 min read",hero:N,heroAlt:"A small figure confronting a towering abstract machine in a field of blue."},l={slug:"nvidia-nemotron-prospect-ranking",category:"Engineering",title:"Why NextLM chose an open NVIDIA Nemotron model for customer-specific prospect ranking.",description:"How NextLM fine-tunes NVIDIA Nemotron models, trains them on NVIDIA DGX Spark, and serves one custom-tuned model per customer on NVIDIA A100 GPUs on Google Cloud.",author:"NextLM · A NextLM engineering story",authorRole:"Engineering",publishedAt:"2026-08-26",publishedLabel:"August 26, 2026",readTime:"14 min read",hero:j,heroAlt:"An abstract blue and violet gradient representing a compact, sparsely activated model architecture."},Y={[o.slug]:o,[l.slug]:l},H=[{slug:o.slug,tag:o.category,title:"Your pipeline has a Goliath problem",description:"Why bigger systems keep missing the few people who are actually ready to move.",image:N},{slug:l.slug,tag:l.category,title:"Why NextLM chose an open NVIDIA Nemotron model",description:"Fine-tuning, training on NVIDIA DGX Spark, and serving one model per customer on NVIDIA A100 GPUs.",image:j}],Q=[{tag:"Teardown",title:"Why account-level intent keeps missing the buyer",description:"The gap between a warm logo and the person who actually decides.",image:X},{tag:"Playbook",title:"What to do in the first hour with an in-market prospect",description:"A simple loop from an in-market prospect to a booked meeting.",image:$},{tag:"Signals",title:"The behavior signals that actually predict a deal",description:"Which triggers matter, and which ones are just noise.",image:q},{tag:"Teardown",title:"The real cost of a cold lead list",description:"What you pay for volume, and what it actually returns.",image:F},{tag:"GTM",title:"Timing the enterprise buy without guessing",description:"When evaluations start, and how to arrive at the right moment.",image:_},{tag:"Product",title:"How your agent learns who you actually close",description:"A plain look at the learning loop behind the score.",image:K}];function J(){return e.jsxs("div",{className:"article-copy",children:[e.jsx("p",{className:"article-lede",children:"Most revenue teams are not short on data. They are surrounded by it: account scores, page visits, keyword surges, enrichment fields, engagement histories, and lists that grow faster than anyone can work them."}),e.jsx("p",{children:"That abundance has created a strange disadvantage. The bigger the system gets, the easier it is to mistake coverage for clarity. Thousands of accounts can look vaguely interesting while the handful of people with a real reason to act disappear into the average."}),e.jsx("h2",{id:"scale-is-not-signal",children:"Scale is not the same as signal."}),e.jsx("p",{children:"Traditional prospecting systems begin with the largest useful object: the account. They collect every observable event around it, roll those events into a score, and ask a seller to find the person behind the number."}),e.jsx("p",{children:"But companies do not research, compare, persuade, or buy. People do. A warm account can contain a cold buying committee. A quiet account can contain one operator already building the case for change."}),e.jsx("blockquote",{children:"The point is not to know more about the market. It is to notice the person whose situation just changed."}),e.jsxs("section",{className:"article-shift","aria-labelledby":"article-shift-heading",children:[e.jsxs("div",{className:"article-shift-heading",children:[e.jsx("p",{children:"Signal model"}),e.jsx("h2",{id:"article-shift-heading",children:"The shift from volume to timing."})]}),e.jsx("div",{className:"article-shift-grid",children:[["01","Account heat","A broad clue that something may be happening."],["02","Person trigger","The individual, responsibility, and change behind the activity."],["03","Relevant action","A next step grounded in why this person might move now."]].map(([t,s,n])=>e.jsxs("div",{className:"article-shift-step",children:[e.jsx("span",{children:t}),e.jsx("h3",{children:s}),e.jsx("p",{children:n})]},t))})]}),e.jsx("h2",{id:"the-precise-advantage",children:"The advantage belongs to the precise."}),e.jsx("p",{children:"David did not beat Goliath by becoming larger. He won by changing the unit of competition. Revenue teams can do the same by replacing broad account awareness with person-level timing."}),e.jsxs("ol",{className:"article-steps",children:[e.jsxs("li",{children:[e.jsx("span",{children:"01"}),e.jsxs("div",{children:[e.jsx("h3",{children:"Start with the person."}),e.jsx("p",{children:"Identify the operator, champion, or decision-maker whose role and current responsibility make the problem theirs to solve."})]})]}),e.jsxs("li",{children:[e.jsx("span",{children:"02"}),e.jsxs("div",{children:[e.jsx("h3",{children:"Look for a change in state."}),e.jsx("p",{children:"A new mandate, a role change, a new tool, repeated research, or a public operational constraint is more useful than undifferentiated activity."})]})]}),e.jsxs("li",{children:[e.jsx("span",{children:"03"}),e.jsxs("div",{children:[e.jsx("h3",{children:"Reach them with the reason now."}),e.jsx("p",{children:"Good outreach connects what changed to what the person must do next. It does not recite everything your database knows."})]})]})]}),e.jsx("h2",{id:"build-for-the-window",children:"Build for the buying window."}),e.jsx("p",{children:"A useful prospecting system should help a team answer three questions in order: Is this the right person? Is there credible evidence that their situation has changed? What should we do next?"}),e.jsx("p",{children:"When the system cannot answer those questions, more volume only makes the uncertainty larger. When it can, a smaller queue becomes an advantage: fewer false positives, more relevant conversations, and more time spent where a deal can actually begin."}),e.jsxs("div",{className:"article-takeaway",children:[e.jsxs("p",{className:"article-takeaway-label",children:[e.jsx(w,{size:15,"aria-hidden":"true"})," The takeaway"]}),e.jsx("p",{children:"Do not outspend the giant on data. Change the unit of attention from accounts that might care to people with a reason to move now."})]})]})}function Z(){return e.jsxs("div",{className:"article-copy",children:[e.jsx("p",{className:"article-lede",children:"NextLM builds an AI prospecting agent for sales teams: one custom-tuned model per customer, trained on that customer's own closed deals. We observe roughly 35 billion behavioral signals a day across the open web and tie them to a named individual inside our identity graph covering 300 million companies and 370 million professional profiles. With this data in hand, sales teams can identify customers who have strong intent to buy, or those who are actively researching their product or service."}),e.jsx("p",{children:"Seeing the behavior is not the hard part. The prioritization is. Of the hundred names in front of a rep on a given morning, which ten are worth the call? Reps do not have a data problem. They have a prioritization problem, and only the top of the list ever gets worked."}),e.jsx("p",{children:"One way to build that ranker is a prompt pointed at a general-purpose frontier model. It launches fast, and it gives up everything downstream: no training on a customer's own won and lost deals, no control over the serving stack, no control over cost. At frontier prices it has to be one shared model for everybody."}),e.jsx("p",{children:"We chose the opposite approach: a small open model we could fine-tune, self-host, and adapt separately for each customer. That model is NVIDIA Nemotron 3.5 Lightning, part of the Nemotron 3 open model family, as a supporting signal that sharpens the ranking through a bounded adjustment. We put the NVIDIA-based system against a broad set of general-purpose frontier language models in a blinded benchmark, each scored zero-shot and with eight examples."}),e.jsx("p",{children:"We expected the much larger frontier models to outperform the smaller open model. They did not."}),e.jsx("h2",{id:"metrics-that-matter",children:"Metrics that matter"}),e.jsx("p",{children:"We benchmark the harder version of our own system on purpose. The product ships a custom-tuned model per customer, but a per-customer model gets to specialize on the exact list it is graded on, and that flatters the number. So for the test we scored the portable, all-business configuration instead: one model, held out on businesses it never trained on. It is the conservative thing to measure, and it is the number we will defend."}),e.jsx("p",{children:"On a blinded audit of 36,704 prospects across 18 businesses, the NVIDIA Nemotron-based system captured 24.6% of all eventual buyers in the top-scoring decile, at a 2.45× lift. Across the field of frontier language models we tested, in both prompting conditions, the closest reached 22.8% and 2.27× on the same test. The NVIDIA Nemotron-based system led the field on all four measures we tracked: top-decile capture and lift for the operating boundary, plus ROC-AUC and PR-AUC for overall ranking quality."}),e.jsx("blockquote",{children:"It's important to say clearly that this was an internal, task-specific benchmark, not a general comparison of model intelligence."}),e.jsx("p",{children:"Frontier models are built for breadth and remain extraordinary at what they are built for. For our narrow prospect-ranking workflow, a specialized open model, fine-tuned on the actual task and served on NVIDIA GPUs, landed higher at the top of the list."}),e.jsx("h2",{id:"top-of-the-list",children:"Why the top of the list is the number that matters"}),e.jsx("p",{children:"Ranking the whole list well is a different question from ranking the top of it well, and revenue teams are paid on the second. A rep has a fixed number of hours and can only touch a boun

1ded number of prospects, so the team works from the top of the list down until capacity runs out, and the long tail is never called no matter how it is ordered. What translates into booked meetings is how many real buyers sit above the line the team can actually reach, which is exactly top-decile capture."}),e.jsx("p",{children:"A model that orders the full distribution slightly better but packs fewer buyers into the reachable top loses the only race that pays. This is why we lead with capture and lift. On this cohort the NVIDIA Nemotron based system also posted the highest ROC-AUC and PR-AUC in the field, so the top-decile lead does not come at the expense of overall ranking, but the top of the list is where the result is put to work."}),e.jsx("h2",{id:"cost-gap",children:"The cost gap is measured, and it is enormous"}),e.jsx("p",{children:"The ranking result stands on its own. The cost gap is the other half of the story, and it is a large, directly measured number."}),e.jsx("p",{children:"On the same 36,704 prospects, our system scored the corpus on a single owned GPU for roughly $0.003 to $0.011 per thousand scores. General-purpose frontier language model APIs on the same job ran between $0.26 and $5.11 per thousand, adding up to over a thousand dollars in metered spend across the field. That is more than an order of magnitude below even the cheapest frontier API, and hundreds to over a thousand times below the most expensive, on hardware we own rather than a meter we rent."}),e.jsx("p",{children:"Sparse activation is the reason a model per customer pencils out at all. A frontier-scale model per customer does not pencil at any volume that matters. A compact NVIDIA Nemotron model with only a small fraction of its parameters active per token, self-hosted, with no per-token cost, does. When your architecture assumes a separate model for every customer, the meter is what decides whether the business can exist."}),e.jsx("h2",{id:"classical-baseline",children:"The classical baseline, too"}),e.jsx("p",{children:"Large language models are not the only comparison a propensity system owes a reader. The workhorse for scoring problems like this one, for over a decade, is the gradient-boosted tree, and a skeptic should ask why a fine-tuned model is needed over a well-tuned XGBoost. Under the same held-out, cross-business protocol on the earlier version of the benchmark where we ran the full baseline suite, the NVIDIA Nemotron based system led there as well: roughly twenty percent more buyers captured in the worked decile than the strongest tree (20.4% versus 17.0%), and a wider gap on overall ranking, where tuned trees cluster near random (0.489 to 0.523 ROC-AUC) once they have to generalize to businesses they never saw, while the Nemotron-based system reached 0.610."}),e.jsx("h2",{id:"why-nvidia-nemotron",children:"Why NVIDIA Nemotron"}),e.jsx("p",{children:"We chose NVIDIA Nemotron for four reasons, in the order they mattered."}),e.jsxs("ol",{className:"article-steps",children:[e.jsxs("li",{children:[e.jsx("span",{children:"01"}),e.jsxs("div",{children:[e.jsx("h3",{children:"The architecture is designed to be small yet mighty."}),e.jsx("p",{children:"The NVIDIA Nemotron family includes compact, sparsely activated models. Our core scorer is fine-tuned from a NVIDIA Nemotron model of roughly 30 billion total parameters (30B A3B), of which only about 3 billion are active for any one input, so compute per score reflects the small active portion rather than the full model size. NVIDIA's newest one, released in August 2026 as NVIDIA Nemotron 3.5 Lightning, extends that efficiency with a hybrid Mamba-2, mixture-of-experts, and attention architecture and a 1M-token context window, and runs on a single GPU. We reached for it as an additional evidence signal that sharpens the ranking through a bounded correction: the correction is cross-validated on held-out businesses and can move any prospect's rank by at most two percentile points, so Lightning enriches the score but never overrides the fine-tune. The Mamba-2 layers handle long behavioral histories at low latency, which lets us feed a full activity sequence rather than a summary of one."})]})]}),e.jsxs("li",{children:[e.jsx("span",{children:"02"}),e.jsxs("div",{children:[e.jsx("h3",{children:"It is open in the dimensions our architecture depends on."}),e.jsx("p",{children:"NVIDIA publishes the weights, substantial training data and recipes, and a commercial-use license. Open weights alone would not have been enough. We needed to own and version the fine-tuned artifact rather than depend on a rented endpoint, and a license that survives contact with an enterprise legal team. We fine-tune with NVIDIA NeMo, specifically NeMo AutoModel, which lets us go from a base NVIDIA Nemotron checkpoint to a customer-specific fine-tune inside our own environment, on NVIDIA DGX Spark, without a separate training pipeline to maintain."})]})]}),e.jsxs("li",{children:[e.jsx("span",{children:"03"}),e.jsxs("div",{children:[e.jsx("h3",{children:"It is a U.S. model developed by a company our customers already have on their approved vendor list."}),e.jsx("p",{children:"That sounds like a procurement detail. In our market it is a gating one. We sell into financial services, staffing, and payroll, and those buyers ask what model is being used, where it was developed, how it was trained, and where customer data goes. With NVIDIA Nemotron we answer all four and point at a published model card and license. That is a short meeting instead of a long one, and a short meeting is sometimes the whole deal."})]})]}),e.jsxs("li",{children:[e.jsx("span",{children:"04"}),e.jsxs("div",{children:[e.jsx("h3",{children:"The serving controls are explicit rather than prose."}),e.jsx("p",{children:"The model exposes reasoning-on and reasoning-off behavior and explicit reasoning budgets through its chat template, so we set behavior per call rather than asking the model in prose not to reason. Structured extraction favors reasoning off with constrained output; engagement analysis benefits from reasoning on."})]})]})]}),e.jsx("h2",{id:"why-dgx-spark",children:"Why we train on NVIDIA DGX Spark"}),e.jsx("p",{children:"We bought two NVIDIA DGX Spark systems rather than making managed cloud fine-tuning our default. Three reasons."}),e.jsxs("section",{className:"article-shift","aria-labelledby":"article-shift-heading",children:[e.jsxs("div",{className:"article-shift-heading",children:[e.jsx("p",{children:"Local training"}),e.jsx("h2",{id:"article-shift-heading",children:"Three reasons we own the hardware."})]}),e.jsx("div",{className:"article-shift-grid",children:[["01","Control of raw customer data","Training inside our own environment means raw closed-won records never get staged with a third-party service. That is a one-sentence answer for a security reviewer."],["02","Cost structure","Training recurs as customer models refresh, so runs scale with customers times time. Owned capacity is a fixed cost, and a model per customer only pencils out with a fixed-cost structure."],["03","Iteration speed","No instance to provision and no dataset to stage before each experiment. Wall clock on an unattended refresh was never our binding constraint. Queue time and data staging were."]].map(([t,s,n])=>e.jsxs("div",{className:"article-shift-step",children:[e.jsx("span",{children:t}),e.jsx("h3",{children:s}),e.jsx("p",{children:n})]},t))})]}),e.jsx("p",{children:"Each NVIDIA DGX Spark pairs an NVIDIA GB10 Grace Blackwell Superchip with 128GB of coherent unified memory and a 200Gb/s NVIDIA ConnectX-7 interface; two of them give a 256GB aggregate for distributed jobs, draw from a conventional outlet, and sit on a desk. The most valuable property is continuity: NVIDIA DGX Spark runs the same CUDA and container stack we use in deployment. The silicon changes between local training and cloud inference. The model format and the serving interface do not."}),e.jsx("h2",{id:"train-serve",children:"Train locally. Serve globally. O

1ne NVIDIA AI stack."}),e.jsx("p",{children:"The training examples are built from real, sealed conversion outcomes, assembled into timestamped features and outcome labels in BigQuery and put in time order before anything reaches a GPU. The expensive part of this work is not the fine-tune. It is producing examples worth training on."}),e.jsx("p",{children:"Training runs locally on our NVIDIA DGX Spark systems. For serving, a shared, quantized NVIDIA Nemotron base model runs on A2 VMs powered by NVIDIA A100 GPUs on Google Cloud, and each customer is assigned a separate LoRA adapter: a small set of additional weights, trained only on that customer's labeled outcomes, to adapt the shared model to that customer's ranking task. The base model is shared. The customer-specific training data, adapters, and access controls are not. Adding a customer adds an adapter, not automatically another GPU. The adapter is only one layer of tenant isolation. Authentication, tenant-aware adapter routing, storage permissions, encryption, logging, and cross-tenant regression tests enforce the rest."}),e.jsx("h2",{id:"gemini-enterprise",children:"Meeting sellers where they already work: native in Gemini Enterprise, on Google Cloud Marketplace"}),e.jsx("p",{children:"A ranking is only useful if it reaches the rep. Rather than build a standalone application that asks a sales team to log into one more portal, NextLM ships natively inside Google Cloud's Gemini Enterprise, among the first agents listed in the Gemini Enterprise agent gallery. NextLM is built on the Agent2Agent (A2A) framework, the open protocol for agent-to-agent communication that Gemini Enterprise uses to orchestrate agents together, so a team works NextLM inside the tool they already have open and named prospects flow through A2A into their CRM alongside the rest of the team's daily agents. An admin turns it on with a single click. There is no browser extension to install, no separate intelligence portal to maintain, and no context-switching to see the score."}),e.jsx("p",{children:"NextLM is also available through Google Cloud Marketplace, so a customer already spending on Google Cloud can procure NextLM through a budget and vendor process already approved. That turns what would otherwise be a new-vendor security and finance review into a line item on an existing contract, which for a large enterprise buyer is often the difference between a six-month sales cycle and a six-week one."}),e.jsx("p",{children:"Together the two integrations mean NextLM is not merely built on Google Cloud, it is delivered through the same surfaces Google Cloud customers already use every day, powered by NVIDIA accelerated computing."}),e.jsx("p",{children:"What lands in front of a rep is an intel report: the named person, the observable behavior that raised the score, and an engagement recommendation. With the fine-tuned NVIDIA Nemotron model, the confidence score is no longer a number the rep has to take on faith. It is a calibrated probability of conversion, grounded in that customer's own historical outcomes, that a sales rep can actually trust."}),e.jsxs("p",{children:["NextLM is a member of ",e.jsx("a",{href:"https://www.nvidia.com/en-us/startups/showcase/?ncid=ref-inc-812304-vt53",target:"_blank",rel:"noopener noreferrer",children:"NVIDIA Inception"}),", a free program that helps AI startups grow through access to the NVIDIA platform, technical guidance, and go-to-market support."]}),e.jsx("h2",{id:"the-takeaway",children:"The Takeaway"}),e.jsx("p",{children:"The overarching takeaway here is not that an open model is universally better than a general-purpose frontier model. It is that a customer-specific ranking problem requires a model architecture that can learn from each customer's own outcomes. A general-purpose frontier model is built to be a generalist. NextLM's product (i.e., ranking a hundred names against one company's closed deals) is not trying to solve every knowledge task; it has one specific job, and on that job, calibrated customer-specific prediction matters more than general knowledge."}),e.jsx("p",{children:"A frontier language model does remain a part of our production system on Google Cloud, doing what it is best at. The fine-tuned NVIDIA Nemotron model produces the score and the structured evidence behind it. The frontier model writes the human-facing report, constrained to explain the score and evidence it receives, not to revise the ranking. Two models running together in a single cloud: a customer-specific, fine-tuned NVIDIA Nemotron model for prediction and calibration; a general-purpose frontier model for communicating the result clearly to a seller."}),e.jsxs("div",{className:"article-takeaway",children:[e.jsxs("p",{className:"article-takeaway-label",children:[e.jsx(w,{size:15,"aria-hidden":"true"})," The takeaway"]}),e.jsx("p",{children:"NVIDIA's open models and NVIDIA DGX Spark systems make customer-specific fine-tuning practical for us, and NVIDIA AI infrastru

1cture on Google Cloud enables us to run our solution at production scale. Google Cloud's Gemini Enterprise puts the result in front of a seller, inside a team's existing workflow. For a task that decides which ten prospects get attention first, that combination of open model, local fine-tuning, and production-grade NVIDIA infrastructure is what makes the whole system possible."})]}),e.jsxs("p",{children:["Discover NextLM on ",e.jsx("a",{href:"https://cloud.google.com/find-a-partner/marketplace",target:"_blank",rel:"noopener noreferrer",children:"Google Cloud's Gemini Enterprise"}),"."]}),e.jsxs("div",{className:"article-note",children:[e.jsx("p",{className:"article-note-label",children:"Technical note"}),e.jsx("p",{children:"The benchmark figures in this article are NextLM internal results, produced by a seeded evaluation harness run on Google Cloud. They compare a task-specific NVIDIA Nemotron system, in its portable all-business configuration held out on unseen businesses, against a broad set of general-purpose frontier language models, each run zero-shot and eight-shot, on NextLM's prospect-ranking dataset. At eleven evaluable businesses the top-decile margins are positive but within their confidence intervals, which the accompanying technical paper reports in full. The cost comparison and the operating-boundary lead are the load-bearing claims. These results are specific to this task and should not be generalized to unrelated workloads."}),e.jsx("p",{children:"This is NextLM's independent work; nothing here should be read as an endorsement by, or a joint claim with, any third party."}),e.jsx("p",{className:"article-fineprint",children:"NVIDIA, NVIDIA Nemotron, NVIDIA NeMo, NVIDIA DGX Spark, NVIDIA Grace Blackwell, NVIDIA GB10, NVIDIA ConnectX, NVIDIA A100, and NVIDIA Inception are trademarks or registered trademarks of NVIDIA Corporation in the United States and other countries. Google Cloud, Gemini, Gemini Enterprise, Google Cloud Marketplace, and BigQuery are trademarks of Google LLC. Agent2Agent (A2A) is an open agent-interoperability framework. All other product names are the property of their respective owners."})]})]})}function ee(){return e.jsxs("section",{className:"article-key-takeaways","aria-labelledby":"key-takeaways-heading",children:[e.jsx("p",{className:"article-section-kicker",children:"The short version"}),e.jsx("h2",{id:"key-takeaways-heading",children:"Key takeaways"}),e.jsxs("ul",{children:[e.jsx("li",{children:"More account data can increase coverage while making the actual buyer harder to see."}),e.jsx("li",{children:"Person-level changes are more useful than broad account activity because they explain who may move and why now."}),e.jsx("li",{children:"The strongest prospecting systems turn fit, timing, and context into a small queue with a clear next action."})]})]})}function te(){return e.jsxs("section",{className:"article-key-takeaways","aria-labelledby":"key-takeaways-heading",children:[e.jsx("p",{className:"article-section-kicker",children:"The short version"}),e.jsx("h2",{id:"key-takeaways-heading",children:"Key takeaways"}),e.jsxs("ul",{children:[e.jsx("li",{children:"On a blinded, held-out benchmark of 36,704 prospects across 18 businesses, a fine-tuned NVIDIA Nemotron model captured 24.6% of eventual buyers in the top decile (2.45× lift), leading every general-purpose frontier model tested, which topped out at 22.8% (2.27× lift)."}),e.jsx("li",{children:"The same system scores prospects for roughly $0.003 to $0.011 per thousand on owned GPUs, versus $0.26 to $5.11 per thousand for frontier language model APIs on the identical job."}),e.jsx("li",{children:"NextLM fine-tunes a sparsely activated NVIDIA Nemotron model (30B total parameters, about 3B active) with NVIDIA NeMo on two NVIDIA DGX Spark systems, then serves one LoRA adapter per customer on NVIDIA A100 GPUs on Google Cloud."}),e.jsx("li",{children:"NVIDIA Nemotron was chosen for its sparse architecture, open license, U.S. provenance, and explicit reasoning controls; a general-purpose frontier model still writes the human-facing report."}),e.jsx("li",{children:"NextLM ships natively inside Google Cloud's Gemini Enterprise and Google Cloud Marketplace, powered by NVIDIA accelerated computing."})]})]})}const ae={[o.slug]:ee,[l.slug]:te},se={[o.slug]:J,[l.slug]:Z},ne={[o.slug]:[{id:"key-takeaways-heading",label:"Key takeaways"},{id:"scale-is-not-signal",label:"Scale vs. signal"},{id:"the-precise-advantage",label:"The precise advantage"},{id:"build-for-the-window",label:"The buying window"}],[l.slug]:[{id:"key-takeaways-heading",label:"Key takeaways"},{id:"metrics-that-matter",label:"Metrics that matter"},{id:"why-nvidia-nemotron",label:"Why NVIDIA Nemotron"},{id:"why-dgx-spark",label:"Why NVIDIA DGX Spark"},{id:"the-takeaway",label:"The takeaway"}]};function re({article:t,shareStatus:s,onShare:n,onUnavailable:c}){const a=ne[t.slug]??[],u=i=>

1{i.preventDefault(),c("Newsletter signup is coming soon.")};return e.jsxs("aside",{className:"article-rail",children:[e.jsxs("nav",{"aria-label":"Article contents",children:[e.jsx("p",{className:"article-rail-title",children:"Table of contents"}),e.jsx("ol",{children:a.map((i,p)=>e.jsx("li",{children:e.jsxs("a",{href:`#${i.id}`,children:[e.jsx("span",{children:String(p+1).padStart(2,"0")})," ",i.label]})},i.id))})]}),e.jsxs("div",{className:"article-rail-share",children:[e.jsxs("button",{type:"button",onClick:n,children:[e.jsx(B,{size:16,"aria-hidden":"true"})," Share article"]}),e.jsx("p",{role:"status","aria-live":"polite",children:s})]}),e.jsxs("div",{className:"article-author-card",children:[e.jsx("p",{children:"Written by"}),e.jsx("h2",{children:t.author}),e.jsx("span",{children:t.authorRole}),e.jsxs("span",{className:"article-author-meta",children:[e.jsx("time",{dateTime:t.publishedAt,children:t.publishedLabel})," · ",t.readTime]})]}),e.jsxs("section",{className:"article-newsletter","aria-labelledby":"article-newsletter-heading",children:[e.jsx("p",{className:"article-newsletter-kicker",children:"The newsletter"}),e.jsx("h2",{id:"article-newsletter-heading",children:"The 3% Rule"}),e.jsx("p",{className:"article-newsletter-body",children:"One sharp teardown a week on the 3% of the market that is actually ready to buy, and how the strongest teams reach them first."}),e.jsxs("form",{onSubmit:u,children:[e.jsx("label",{htmlFor:"article-newsletter-email",className:"sr-only",children:"Work email"}),e.jsx("input",{id:"article-newsletter-email",type:"email",required:!0,autoComplete:"email",placeholder:"Enter your email"}),e.jsxs("button",{type:"submit",children:["Subscribe ",e.jsx(v,{size:16,"aria-hidden":"true"})]})]}),e.jsx("p",{className:"article-newsletter-fineprint",children:"Weekly. Unsubscribe anytime."})]})]})}function ie({currentSlug:t}){const s=H.filter(a=>a.slug!==t).map(a=>({tag:a.tag,title:a.title,description:a.description,image:a.image,href:`/articles/${a.slug}`,action:"Read article"})),n=Q.map(a=>({...a,href:"/articles",action:"Read soon"})),c=[...s,...n].slice(0,6);return e.jsxs("section",{className:"article-related","aria-labelledby":"related-articles-heading",children:[e.jsxs("div",{className:"article-related-heading",children:[e.jsx("p",{className:"article-section-kicker",children:"Keep reading"}),e.jsx("h2",{id:"related-articles-heading",children:"More from NextLM"})]}),e.jsx("div",{className:"article-related-grid",children:c.map(a=>e.jsxs("article",{children:[e.jsxs("div",{className:"article-related-image",children:[e.jsx("img",{src:a.image,alt:"",loading:"lazy"}),e.jsx("span",{children:a.tag})]}),e.jsxs("div",{className:"article-related-copy",children:[e.jsx("h3",{children:a.title}),e.jsx("span",{children:a.description}),e.jsxs("a",{href:a.href,children:[a.action," ",e.jsx(v,{size:12,"aria-hidden":"true"})]})]})]},a.title))})]})}function oe({onUnavailable:t}){return e.jsx("section",{className:"article-closing",style:{backgroundImage:`url(${z})`},"aria-labelledby":"article-closing-heading",children:e.jsxs("div",{children:[e.jsx("p",{children:"Keep going"}),e.jsx("h2",{id:"article-closing-heading",children:"Everything here points somewhere useful."}),e.jsxs("div",{className:"article-closing-links",children:[e.jsxs("a",{href:"/events",children:[e.jsx(U,{size:19,"aria-hidden":"true"})," Join a live workshop"]}),e.jsxs("button",{type:"button",onClick:()=>t("Recordings are coming soon."),children:[e.jsx(E,{size:19,"aria-hidden":"true"})," Watch recordings on YouTube"]})]})]})})}function le(){return e.jsx("main",{className:"flex min-h-[62vh] items-center justify-center px-6 py-24 text-center",children:e.jsxs("div",{className:"max-w-xl",children:[e.jsx("p",{className:"text-[11px] font-semibold uppercase tracking-[0.2em]",style:{color:y},children:"Article not found"}),e.jsx("h1",{className:"mt-5 text-4xl font-medium text-[#171717] sm:text-5xl",children:"This story is not here yet."}),e.jsx("p",{className:"mt-5 text-[16px] leading-7 text-[#656565]",children:"Browse the latest thinking on buyer timing, signals, and modern prospecting."}

1),e.jsxs("a",{href:"/articles",className:"mt-8 inline-flex items-center gap-2 rounded px-5 py-3 text-[14px] font-medium text-white",style:{background:y},children:[e.jsx(S,{size:16,"aria-hidden":"true"})," Back to articles"]})]})})}function be(){const{slug:t=""}=D(),s=Y[t],n=ae[t],c=se[t],a=G(),{blocks:u,loading:i}=V(),[p,m]=h.useState(""),[g,f]=h.useState(""),[x,k]=h.useState(0),I=u.user.source==="live"||u.user.source==="demo";h.useEffect(()=>{const r=document.documentElement.style.overflow;return document.documentElement.style.overflow="auto",document.title=s?`${s.title} | NextLM`:"Article not found | NextLM",()=>{document.documentElement.style.overflow=r}},[s]),h.useEffect(()=>{const r=document.documentElement,d=r.classList.contains("dark");return r.classList.remove("dark"),()=>{d&&r.classList.add("dark")}},[]);const A=async()=>{if(s)try{navigator.share?(await navigator.share({title:s.title,text:s.description,url:window.location.href}),m("Article shared.")):(await navigator.clipboard.writeText(window.location.href),m("Link copied."))}catch(r){if(r instanceof DOMException&&r.name==="AbortError")return;m("Could not copy the link.")}},T=r=>{const d=r.currentTarget,b=d.scrollHeight-d.clientHeight;k(b>0?Math.min(1,d.scrollTop/b):0)};return M.createPortal(e.jsx("div",{onScroll:T,className:"nextlm-light-page fixed inset-0 z-[9999] overflow-y-auto bg-white",style:{fontFamily:R,scrollBehavior:"smooth"},children:e.jsxs("div",{className:"min-h-screen text-[#171717]",children:[e.jsx(W,{activePath:"/articles",isAuthed:I,isLoading:i,onLogin:C,onStart:L,onDashboard:()=>a("/dashboard")}),e.jsx("div",{className:"sticky top-[69px] z-40 h-[3px] w-full overflow-hidden bg-[#e7e7e7]",role:"progressbar","aria-label":"Article reading progress","aria-valuemin":0,"aria-valuemax":100,"aria-valuenow":Math.round(x*100),children:e.jsx("span",{className:"block h-full origin-left bg-[#306AFF] transition-transform duration-150 motion-reduce:transition-none",style:{transform:`scaleX(${x})`}})}),s?e.jsx("main",{className:"bg-white",children:e.jsxs("article",{children:[e.jsx("div",{className:"article-band",style:{backgroundImage:`url(${O})`},children:e.jsx("span",{children:"Articles"})}),e.jsxs("div",{className:"article-reading-surface",children:[e.jsxs("div",{className:"article-reading-grid",children:[e.jsxs("div",{className:"article-reading-column",children:[e.jsx("figure",{className:"article-hero-art",children:e.jsx("img",{src:s.hero,alt:s.heroAlt,loading:"eager"})}),e.jsx("h1",{className:"article-title",children:s.title}),e.jsx("p",{className:"article-description",children:s.description}),n&&e.jsx(n,{}),c&&e.jsx(c,{})]}),e.jsx(re,{article:s,shareStatus:p,onShare:()=>void A(),onUnavailable:f})]}),e.jsx(ie,{currentSlug:t})]}),e.jsx(oe,{onUnavailable:f})]})}):e.jsx(le,{}),e.jsx("p",{className:"sr-only",role:"status","aria-live":"polite",children:g}),g&&e.jsxs("div",{className:"fixed bottom-6 left-1/2 z-[10000] -translate-x-1/2 rounded bg-[#212121] px-4 py-3 text-[14px] text-white",children:[g,e.jsx("button",{type:"button",className:"ml-4 text-[#bcfc00]",onClick:()=>f(""),children:"Dismiss"})]}),e.jsx(P,{})]})}),document.body)}export{be as default};