Most GPU decisions start in the wrong place.
They start with a product page, a familiar accelerator name, or a cloud instance someone used on the last project. Then comes the spreadsheet: model weights, context length, KV cache, concurrency, traffic, headroom, hourly rates, provider availability. One uncertain input quietly becomes five confident-looking numbers.
By the time a team gets to a recommendation, it can be difficult to tell whether the configuration fits the workload—or whether the spreadsheet was built to justify the configuration.
The new Skyportal AI Compute Advisor reverses that process.
Start with the workload. Skyportal turns a plain-English description into a complete compute plan: the GPU configuration it recommends, why it fits, what it should cost, which alternatives were evaluated, where capacity is available, and which assumptions could change the answer.
No account is required for the first recommendation. No spreadsheet is required at all.
Describe the system you are building—not the instance you think you need
The Advisor begins with the information engineering teams already have:
We're serving Llama 3.3 70B for a support copilot. Around 50,000
requests an hour at peak, prompts about 4,000 tokens. Currently on
8× H100 at $2.90/GPU/hr.
That description contains several different planning signals: model size, traffic, traffic basis, prompt length, current infrastructure, and current cost. Skyportal normalizes them into a workload specification and fills only the gaps it needs to produce an answer.
You can also begin with less. Enter a model name to see its memory requirements across supported precisions. Enter a monthly budget to explore what fits. Describe an existing deployment to find out whether it is appropriately sized—or whether you are paying for capacity the workload cannot use.
The interface adapts to the question instead of forcing every user through the same intake form.
One recommendation, with the decision exposed
The result is not a wall of instance types. It is a decision page.
At the top, the Advisor makes the call: a recommended GPU and count, estimated monthly compute cost, capacity headroom, and a plain-language verdict. Below that, it shows the work behind the answer:
- Every viable configuration evaluated on the same workload
- Memory fit and usable VRAM
- Aggregate memory-bandwidth proxy
- Target time to first token
- Capacity headroom
- Estimated monthly compute cost
- A clear verdict for each option, from recommended to viable alternative or overkill
If you include the GPUs and rate you pay today, the Advisor keeps that current fleet in the comparison. A cheaper existing setup is not automatically declared a winner: it must also have enough memory to run the workload. If it fits and beats the market alternatives, Skyportal says to stay where you are. If it does not fit, the lower price does not get to hide that fact.
That distinction sounds obvious. It is also exactly the kind of distinction that disappears in a cost-only comparison.
AI for understanding; deterministic logic for the decision
“AI advisor” should not mean “a model guessed a GPU.”
Skyportal can use language understanding to extract workload details from the many ways people describe them. The sizing decision itself then runs through a deterministic evaluation engine. The same normalized workload produces the same memory calculation, evaluates the same catalog, and applies the same selection rules.
The engine accounts for the pieces that dominate inference memory:
- Model weights at the selected precision
- KV cache from context length and concurrent sequences
- Usable GPU memory rather than headline VRAM alone
- Operational headroom before a configuration becomes the primary recommendation
It evaluates supported GPU counts across priced accelerator families, rejects configurations that do not fit, and ranks the remaining options using current cost and capacity data. The result includes its confidence level and exposes the technical breakdown for teams that want to inspect the math or API response.
Natural language makes the tool fast to use. Deterministic sizing makes the recommendation inspectable.
Assumptions are part of the answer
Every compute estimate contains assumptions. Hiding them does not make the estimate more accurate; it only makes it harder to challenge.
When a description leaves out traffic shape, prompt length, response length, latency target, or precision, the Advisor says so directly. Assumptions appear beside the result—not buried in a methodology page—and each is connected to the workload field it affects.
That makes uncertainty actionable. A team can immediately see whether the recommendation depends on an approximate request rate, a default context length, or an inferred peak multiplier. The confidence indicator explains when those gaps affect the plan.
In the Professional Compute Plan, those values become an in-place scenario editor. Change an assumption, apply it, and the recommendation re-runs without throwing away the plan. Staged changes remain visible even when an adjustment requires an upgrade, so the decision context is never lost.
A live market view, not a stale price table
A technically valid GPU plan is not useful if nobody can buy it at the price shown.
The Advisor connects sizing to Skyportal's live pricing catalog. Today that feed spans 24 provider catalogs and thousands of current offers across hyperscalers, specialist GPU clouds, and marketplaces. For the recommended configuration, the result shows matched providers, regions, per-GPU hourly rates, and estimated monthly cost, with the least expensive current match called out.
Pricing freshness is visible. If a live feed is unavailable, Skyportal degrades explicitly instead of presenting fallback data as current market truth. Every estimate remains a compute-only reference, and the result reminds you to validate availability before purchasing.
The catalog can change without changing the sizing engine. That separation matters: model geometry and GPU memory are technical facts; an available instance in a particular region at a particular price is a market fact. They should not be maintained as one hard-coded table.
Built to travel with the decision
A recommendation rarely belongs to one person. It goes to an engineering lead, a finance partner, an infrastructure team, or the person who has to approve the purchase.
Every result can be downloaded as a concise Markdown compute plan. It can also be shared through a durable public report with recommendation-specific preview metadata. Shared reports deliberately omit the caller's original free-text workload description, preserving the decision without exposing the prompt that produced it.
The free report answers the immediate sizing question. The Professional Compute Plan carries that analysis into a deeper planning workspace for architecture, market sourcing, alternatives, sensitivity analysis, and risk assessment.
The goal is a continuous path from “what should we run?” to a plan an engineering team can evaluate and execute.
Who the Compute Advisor is for
The Advisor is built for anyone making an AI infrastructure decision before there is time—or budget—for a full capacity-planning exercise:
- AI product teams moving a model from prototype to production
- Platform engineers comparing GPU clouds and instance shapes
- MLOps teams planning vLLM-style inference deployments
- Finance and engineering leaders validating an existing GPU bill
- Startups deciding what a model can support within a fixed monthly budget
- Teams that know their workload but do not want to become GPU-sizing specialists
It is equally useful as a first answer and as a challenge to an existing one. If your current plan is right, the comparison should make that visible. If it is not, you should be able to see exactly which assumption or constraint changes the decision.
Try it now
Open the Skyportal AI Compute Advisor, describe the model or workload in your own words, and get a recommendation in seconds.
Bring the traffic numbers if you have them. Bring the current GPU bill if you want to challenge it. If important details are missing, Skyportal will show you what it assumed.
The most expensive GPU is not automatically the safest choice. The cheapest instance is not automatically the smartest one. The right compute plan is the one that fits the workload, preserves enough headroom, and makes every important assumption visible.
Now you can get that plan before you commit the spend.