Every act of creation is first an act of destruction
Pablo Picasso
AI models are rapidly evolving from mere parrots to intensive enterprise data processors.
The evolution is so swift that we track progress in weeks rather than years. Consider, for example, the near-immediate transition from LLMs that were master imitators, ingesting and regurgitating the contents of the internet, to full-scale reasoning machines capable of sophisticated chains of thought and human-like creativity.
From a data and infrastructure perspective, these new reasoning models represent a completely new computing paradigm. Everything changes—from training and inference down to the way systems are architected. Why? Because the ways we have built systems decades ago have been informed how they are still designed and deployed today in many cases. But for any organization priming for next-next-gen AI specific to their own business, that cannot continue.
Enterprise data and AI are going to collide and when they do there will be a supernova effect, Jeff Denworth, VAST cofounder told a group of top industry analysts in a call last week. When that happens, enterprise systems designed around limitations and workarounds will be woefully inadequate to the hypercompetitive tasks ahead.
Imagine a Fortune 500 company in retail, for example, which has generated 100 petabytes of structured and unstructured data over its entire history. What if it was possible to vectorize all of that data so it is contextually ready on a system that could be scaled to support trillions of vectors? What would be observed if that was run across a Stargate-class AI system over the course of ten minutes for a 360-degree view of an enterprise’s past, present, and future?
Data Type | % of 100PB | Vectorization Method | Avg. Vector Size (KB) | Vectors per PB | Total Vectors (100PB) |
|---|---|---|---|---|---|
Text (Documents, Logs, PDFs, Code, Emails, etc.) | 40% (40PB) | Chunking (512 tokens) | 1.5 KB | ~660PB | ~26.5T |
Images (Photos, Scans, Satellite, Medical Imaging, etc.) | 20% (20PB) | 1 vector per image (~1MB avg.) | 4 KB | ~250PB | ~5T |
Audio (Meetings, Call Transcripts, Music, etc.) | 10% (10PB) | Speech-to-text → Text Embedding | 2 KB | ~500PB | ~5T |
Video (Surveillance, Media, Presentations, etc.) | 20% (20PB) | Frame sampling (~1 per second) | 4 KB | ~250PB | ~5T |
Scientific Data (Genomics, Simulations, Sensors, etc.) | 10% (10PB) | Domain-specific embeddings | 4 KB | ~250PB | ~2.5T |
(A Fortune 500 company with 100PB of data collected in its history could average five trillion vectors per data type, says VAST co-founder Jeff Denworth.)
While it’s not feasible for most enterprises to have that level of AI hardware at the ready today, the trajectory of scale and capability shows infrastructure demanding mounting fast. It will become the primary point of competition.
Human Brain(1) | 1,024 GPU Machine(2) | 1M GPU Machine | 1M GPU Advantage | |
|---|---|---|---|---|
Cognitive Throughput | 40 MB/s | 160 GB/s* | 320 TB/s* | 8.3Mx |
Storage Capacity | 2.5 Petabytes | 4 Petabytes | 4 Exabytes | 1.6Kx |
FLOPS | 10 ExaFLOPS | 82 PetaFLOPS | 82 ExaFLOPS | 8.2x |
Power | 20 W | 2 mW | 2 gW | -100Mx |
(A glimpse into potential GPU evolution, via Jeff)
We all know what the hyperscalers are doing, at least to some extent. We can all see what’s possible with hundreds of thousands of GPUs but what about the near-trillion-dollar path being carved for million-GPU machines shuttling data at 320 TB/s. And whether it’s large enterprise, the biggest cloud builders, or the world’s top research shops, the key phrase for 2025 should be “future proofing.”
Hedging for the IT future used to be easier, though scalability petered out quickly both on hardware and software fronts. There is no time for that now.
The future of enterprise capability starts with an architecture that can handle data and its manipulation at scale, all without having wait time for results that mount with data volume. It requires an integrated system that becomes the core of a thinking machine at the ready no matter where a user engages with it (at the file system, vector database levels or a runtime for pipeline execution for example).
There is no room for legacy. There is no time for legacy.
Just as it has since its inception, VAST is taking its cues from the largest companies with pressing AI challenges: What do companies need to build enterprise AI pipelines and where else could one find something capable?
The answer, for many thus far, has been to look to the cloud. And while there is capacity and capability on hand, they quickly see that same legacy anchor where there are ten, twenty, thirty or more services that need to be stitched together, none of which are purpose-built. They’re stuck together with glue code and have a square peg/round hole problem, even if they “work.”
Ultimately, whether on-prem or in the cloud, enterprises face tradeoffs that have to be made between database, or storage, or even compute services.
What VAST presents is a reconciliation for the whole explosion of products, technologies, and approaches that are byproducts of the bad architecture that came before.
From building the world’s most scalable and powerful vector database to developing a vision for a thinking machine that keeps all data at the ready, VAST is condensing the capabilities required for next generation enterprise AI into a system that keeps data close, fast, and ready for analysis at massive scale.
This is why every single LLM builder is using VAST. This is why VAST has become the de facto standard at sites that quickly scale from 100 petabytes to exabytes in weeks. And this is the reason for VAST’s unprecedented growth—Jeff, Renen, and team are being summoned to displace legacy systems because they understand old architectures can’t scale to the new levels required.
How is your organization future-proofing for the AI era? When do you think we’ll see a 1m GPU machine? Join the conversation on Cosmos and let us know. And if you’re attending NVIDIA GTC next week in San Jose, visit the VAST booth #733 and learn more about the coming enterprise AI and data supernova.
