Gartner predicts that 40% of generative AI solutions will be multimodal by 2027, up from just 1% in 2023. Multimodal means vision, audio, text, and video running together, and each of those inputs costs far more to process than a line of text. As adoption climbs and per-request compute rises with it, enterprise GPU requirements are set to roughly triple by 2027.
The Shift Is Already Under Way
For most of the generative AI era, enterprise workloads have been dominated by text: chat assistants, summaries, search, and code. That is no longer where the market is heading. Gartner predicts that 40% of generative AI solutions will be multimodal by 2027, up from just 1% in 2023. In other words, the share of AI that handles more than text is rising roughly fortyfold in four years.
The longer trend is just as steep. Gartner also predicts that 80% of enterprise software and applications will be multimodal by 2030, up from less than 10% in 2024. Multimodal is therefore not a niche feature for a few creative tools. It is becoming the default way enterprise software sees, hears, reads, and responds. The reason matters for infrastructure: every one of those new capabilities runs on GPUs, and they do not run cheaply.
Why Multimodal Costs So Much More Per Request
The core issue is data volume. Text is compact, while images, audio, and video are not. Research on vision-language models shows that a single 336-by-336 pixel image becomes 576 vision tokens once the model encodes it. A 672-by-672 image becomes up to 2,880 tokens. A typical text prompt, by contrast, is a few dozen tokens. In common multimodal tasks, vision tokens account for around 80% of the total workload.
Token count alone understates the problem, because attention cost grows with the square of the sequence length. Double the tokens and you roughly quadruple the attention compute. A request that bundles an image, a caption, and a spoken question is therefore far more expensive than the same question in text. Measured energy figures bear this out. Studies find that text-plus-image inference uses 1.1 to 5.2 times the energy per token of text alone. Text-plus-video uses 1.3 to 15 times as much.
Video is the heaviest case of all. A few seconds of generated or analysed video can carry millions of tokens. That is why the most demanding AI workloads of 2026 are visual rather than textual. We examined the economics of inference at scale in AI Inference Costs at Scale.
Four Modalities, One Machine
A modern multimodal request does not run one model. It runs several stages on the same infrastructure: a vision encoder to read the image or video, an audio encoder for speech, and a language model to tie it together. Each stage has a different hardware profile. Vision encoding is compute-bound, while text generation is bound by memory bandwidth.
That mix is hard to serve well. The infrastructure has to deliver high compute and high memory bandwidth at once. It also has to keep several modalities resident and ready rather than swapping models in and out. As enterprises layer vision, audio, text, and video into a single product, the load is not additive so much as compounding. The modalities run concurrently and share the same long context.
Why We Project a Tripling by 2027
Here is how the tripling estimate is built, and where its limits lie. Two verified trends drive it. First, the share of generative AI that is multimodal rises from 1% to 40% by 2027, a roughly fortyfold jump in how often these heavier workloads run. Second, each multimodal request costs several times the compute of a text request, as the token and energy figures above show.
Put those together against a base where inference is already the fastest-growing part of AI demand. Multimodal moves from a rounding error to a large share of all inference, and each of those requests carries multiples of the compute of text. On that basis, the aggregate enterprise GPU requirement for AI is on track to roughly triple by 2027. This is a forward projection rather than a measured number. The exact multiple depends on how fast adoption moves and how much model efficiency improves. The direction, though, is not in doubt.
What This Means for Infrastructure
Planning for multimodal means planning for memory and throughput, not just headline compute. The long token sequences that images and video produce have to fit in GPU memory. The concurrent stages also need bandwidth to move data between the encoders and the language model. High-memory GPUs and fast interconnect matter more here than in a text-only world. The hardware trade-offs across the current generation are set out in our NVIDIA Blackwell GPU comparison.
It also means planning for capacity that scales with a moving target. A team cannot know in 2026 exactly how much multimodal load it will carry in 2027. So the ability to add the right hardware quickly, in the right location, becomes a strategic advantage. That favours dedicated bare-metal capacity with room to grow over fixed, shared instances. The Build program lets enterprises specify GPU type, memory, and fabric and have Axe configure it. For the wider planning picture, see Enterprise GPU Strategy in 2026.
The takeaway is simple. Multimodal AI is not a heavier version of text AI. It is a different workload with a different cost curve, and that curve points sharply upward through 2027. Some enterprises will size their infrastructure for vision, audio, and video now, rather than after the demand arrives. Those are the ones that will ship multimodal products without waiting for capacity.
About Axe Compute
Axe Compute Inc. (NASDAQ: AGPU) is a neocloud AI infrastructure platform built on a fundamental premise: AI innovation should not be constrained by hardware choice or inventory limitations. Axe Compute gives enterprises and AI innovators choice across hardware, geography, and deployment speed through two delivery models: Axe Compute Access, providing the latest GPU compute options in as fast as 48 hours across numerous global locations, and Axe Compute Build, enabling enterprises to access large-scale dedicated AI factories, all backed by enterprise-grade SLAs and support. Axe Compute is headquartered in Pittsburgh, Pennsylvania. For more information, visit axecompute.com.
Size your infrastructure for multimodal, before the demand arrives.
Frequently Asked Questions
Why does multimodal AI need so much more GPU power than text AI?
Because images, audio, and video carry far more data than text. Research on vision-language models shows a single 336-by-336 pixel image becomes 576 vision tokens, and a 672-by-672 image becomes up to 2,880 tokens, where a short text prompt may be only a few dozen tokens. Vision tokens make up around 80% of the workload in common multimodal tasks. Since attention cost scales with the square of sequence length, more tokens raise compute sharply. Measured energy per token is 1.1 to 5.2 times higher for text-plus-image and 1.3 to 15 times higher for text-plus-video.
How fast is multimodal AI being adopted by enterprises?
Very fast. Gartner predicts that 40% of generative AI solutions will be multimodal by 2027, up from just 1% in 2023. Gartner also predicts that 80% of enterprise software and applications will be multimodal by 2030, up from less than 10% in 2024. The share of AI workloads that combine vision, audio, text, and video is rising by roughly 40 times in four years.
Will multimodal AI really triple enterprise GPU requirements by 2027?
That is Axe Compute’s forward projection, not a measured figure. It combines two verified trends: the share of generative AI that is multimodal rising from 1% to 40% by 2027, and the higher per-request compute of multimodal inference, which runs several times the cost of text alone. As multimodal moves from a rounding error to a large share of inference, and each request costs more, aggregate enterprise GPU requirements for AI plausibly triple. The exact multiple depends on adoption speed and model efficiency.
What infrastructure does multimodal AI need?
Multimodal workloads need high-memory GPUs to hold long token sequences, high throughput for vision encoding, and dedicated capacity that stays available as several modalities run at once. That favours bare-metal clusters over shared instances. Axe Compute provides bare-metal GPU infrastructure across 200+ locations in 93 countries, provisioned in approximately 48 hours, with the choice to configure GPU memory and fabric through the Build program.
Sources
- Gartner (September 2024): 40% of generative AI solutions will be multimodal by 2027, up from 1% in 2023
- Gartner (July 2025): 80% of enterprise software and applications will be multimodal by 2030, up from less than 10% in 2024
- “Modality Inflation: Energy Characterization for MLLM Inference,” arXiv (2025): text-plus-image uses 1.1 to 5.2x and text-plus-video 1.3 to 15x the energy per token of text alone
- Research on vision-language model efficiency: a 336×336 image encodes to 576 vision tokens and a 672×672 image to up to 2,880; vision tokens are ~80% of the workload, with quadratic attention cost
- MarketsandMarkets: multimodal AI market growing at roughly 35% CAGR