Asking Three AI Assistants Which GPU Cloud to Use: Perplexity Never Once Mentioned AWS or Google Cloud

More and more developers are asking an AI assistant before they ever open a pricing page: "Which GPU cloud should I use for inference?" This habit is quietly reshaping how cloud services get selected. One researcher ran a curiosity-driven experiment — posing the same questions to ChatGPT, Gemini, and Perplexity to see exactly what each one would recommend.

The results were surprising: ChatGPT and Gemini mostly steered users toward hyperscalers, while Perplexity took a completely different path. In this round of testing, it didn't mention AWS or Google Cloud even once.

How the Experiment Was Designed

The researcher crafted 8 buyer-perspective questions, all free of brand names, covering typical GPU inference selection scenarios:

• Which vendors are good for budget-constrained startups doing prototype validation?
• Which AI inference GPU cloud service is overall the best right now?
• Which providers work well for small teams that want fast onboarding and solid support?
• Which vendors excel at real-time, low-latency inference for end users?
• Which providers offer the best cost efficiency for high-throughput batch inference?
• Which services integrate best with common MLOps toolchains and CI/CD pipelines?
• Which vendors offer enterprise-grade security, compliance, and dedicated GPU options?
• Which providers offer the best hybrid cloud/multi-cloud and on-premises integration?

Each question was sent once to each assistant via their official APIs, using lightweight fast models: gpt-5-mini for ChatGPT, gemini-3.5-flash-lite for Gemini, and pplx-fast for Perplexity. This produced 24 responses in total. A secondary model then parsed each response, extracted the recommended products and their rankings, and tallied everything up.

All three models used in the experiment are lightweight, fast-tier versions from their respective families. gpt-5-mini is OpenAI's small-parameter model optimized for low-latency, low-cost use cases; gemini-3.5-flash-lite is the fastest and cheapest model in the Gemini lineup; and pplx-fast is Perplexity's search-augmented rapid inference model. Choosing these over flagship versions (like GPT-5 or Gemini 2.5 Pro) means the experiment measured what a typical "quick developer query" looks like — not deep-research-mode output quality. Lightweight models tend to have shallower reasoning and narrower knowledge coverage than flagships, making them more likely to rely on high-frequency patterns from training data. This may have amplified the ChatGPT-favors-AWS and Gemini-favors-Google-Cloud effect. The researcher explicitly noted that consumer-facing apps with larger models and browsing capabilities may produce different results.

Three Assistants, Three Preferences

Across the 24 aggregated responses, Google Cloud was mentioned most often at 13 times, followed by AWS and Azure at 10 each, Lambda at 9, RunPod at 8, and CoreWeave, Modal, and Groq in descending order. A total of 20 products were named.

But the really interesting story is in the distribution:

• ChatGPT most frequently recommended AWS — 7 out of 8 responses mentioned it.
• Gemini most favored Google Cloud — 7 out of 8 responses mentioned it.
• Perplexity most often recommended RunPod (4 out of 8 responses) and never mentioned Google Cloud or AWS at all.

In other words, if your team makes decisions by "just asking AI," the answer depends heavily on which AI you ask. This isn't a subtle difference in taste — it's a fundamental divergence across the entire recommendation list.

Most Mentioned ≠ Ranked First

The data reveals a distinction that's easy to overlook: mention frequency and ranking position tell two different stories.

Google Cloud appeared in the most responses (13 times), but was ranked first only twice. AWS was ranked first in 5 of its 10 mentions, with an average rank of 2.0. Groq, despite being mentioned only 3 times, was ranked first every single time.

This finding has real value for product teams tracking AI visibility: mention count measures breadth of exposure, while first-place ranking measures strength of recommendation. You need to watch both — focusing on just one leads to incomplete conclusions.

Another robust signal is Lambda — it was mentioned exactly 3 times by each of ChatGPT, Gemini, and Perplexity, making it the most consistent pick across all three assistants. While it didn't top the charts for any single assistant, that cross-assistant balance suggests broader, more evenly distributed mindshare. By contrast, Azure and RunPod were also mentioned by all three, but with heavily skewed distributions (5/4/1 and 2/2/4, respectively).

Assistants Are Mixing Up Categories

One pattern worth flagging: when asked about "GPU clouds for inference," some responses recommended OpenAI, Anthropic, and Cohere — which sell model APIs, not GPUs — and others recommended Groq, which runs its own LPU hardware rather than traditional GPUs.

The researcher chose to keep these responses in, because "this is what a real buyer would actually read." It also highlights a common failure mode: AI responses often conflate "where I run my own model" with "whose model I'm calling." For developers who aren't familiar with the underlying distinctions, this kind of confusion can directly mislead purchasing decisions.

The underlying taxonomy is worth spelling out. GPU clouds (like Lambda, RunPod, CoreWeave) provide raw compute rental — you deploy your own model weights, manage the inference stack yourself, and pay per GPU-hour. Model API services (like OpenAI, Anthropic, Cohere) are packaged black-box interfaces where you call their hosted models, never touching the underlying hardware, and pay per token. LPUs (Language Processing Units) are Groq's custom inference chips — architecturally distinct from NVIDIA GPUs, with extremely low latency but support for only specific models, used more like an API than traditional GPU rental. These three categories represent completely different technical paths and cost structures: GPU clouds require an engineering team to maintain the inference stack but offer maximum flexibility; model APIs are plug-and-play but can't use private weights; LPUs are somewhere in between. Lumping all three into the same recommendation list is like comparing "buying land to build a warehouse" with "renting a delivery service" — it creates real, practical confusion for developers evaluating their options.

Where Perplexity's Answers Come From

Because Perplexity returns its cited sources, we can see what it actually "read." The most frequently cited sites across 8 responses included gmicloud.ai, io.net, digitalocean.com, and inworld.ai (4 times each), followed by dev.to, northflank.com, siliconflow.com, and spheron.network (3 times each).

Two things stand out:

First, most of these are vendor blogs and comparison articles — not neutral reviews. Whoever wrote the clearest "best GPU cloud for this use case" article is more likely to get read and cited by the assistant.

Second, dev.to was cited 3 times across 8 responses. Developer community content is itself becoming source material for AI-generated GPU recommendations.

This likely explains the divergence: Perplexity answers based on real-time web search results, which are saturated with vertical vendors' own comparison content; ChatGPT and Gemini rely more heavily on training data, where hyperscalers dominate. The researcher acknowledges this is speculation — 24 responses aren't enough to prove it.

This phenomenon has a corresponding concept emerging in the SEO world: GEO (Generative Engine Optimization). Traditional SEO aims to rank pages higher in search results; GEO aims to get content read, cited, and incorporated into AI-generated responses. The core strategies differ: GEO prioritizes content structure (comparison tables AI can easily parse, clear scenario-to-conclusion mappings), authority signals, and precise alignment with query intent — rather than the backlink counts or keyword density that traditional SEO relies on. The rise of real-time search assistants like Perplexity makes GEO especially valuable for niche vendors: the brand recognition advantage built up through training data doesn't apply here, and a single high-quality comparison article published recently can immediately enter the recommendation pool. This also explains why smaller GPU cloud providers appear far more often in Perplexity's recommendations than in ChatGPT's or Gemini's.

Are the Results Stable?

The researcher ran the same questions a few days earlier. The top two were unchanged: Google Cloud (12 responses) and AWS (9 responses). But further down the rankings, things shuffled — RunPod, Modal, Azure, and CoreWeave each moved a few positions.

The conclusion: the broad picture held, but at this sample size, a difference of 1 to 2 responses between vendors should be treated as noise rather than signal.

Key Caveats to Keep in Mind

This experiment has clear boundaries, which the author honestly laid out:

• These are API responses from fast, cheap models. Consumer-facing apps with browsing, memory, and larger models may produce different results.
• 24 responses are a snapshot, not a census. A difference of 1–2 responses should be read as a tie.
• It measures what assistants say, not which provider is actually best. This is not a benchmark for price, latency, or reliability.
• No vendors paid to influence their inclusion, exclusion, or ranking.

What This Means for Developers

The value of this small experiment isn't in delivering a definitive "best GPU cloud" answer. It's in surfacing a reality that's taking shape: as the first step in vendor selection shifts from "open the pricing page" to "ask AI," which AI you use may matter more than what you ask.

For vendors looking to improve their AI visibility, Perplexity's source data offers a clear signal — publishing well-structured, scenario-specific comparison content gets you into recommendation results faster than waiting to accumulate presence in training data. For developers, it's worth asking multiple assistants and cross-checking results, especially when a response conflates model APIs with GPU hardware.

An open question to leave you with: for this kind of vendor selection, do you trust an assistant that searches the web in real time (Perplexity), or one that answers from trained knowledge (ChatGPT, Gemini)?