AI ecosystems in full swing at RAISE Summit – News

I’ve been to several AI-focused conferences since I moved to Europe in 2023. Most of them suffer from trying too hard – driving the hype around AI beyond its current utility.

This week’s RAISE Summit in Paris was a refreshing exception. I spoke with several exhibitors at this packed conference, and the overall buzz centered on the maturation of the AI ecosystem.

Unlike other conferences, none of the vendors was bothering to convince attendees that AI was good for this or that purpose. Everybody now understands its value. Instead, the focus was on providing all the various enabling technologies necessary to achieve that value at scale. In other words, the AI ecosystem is maturing, with each innovative vendor contributing its efforts to enterprise AI strategies of all kinds.

Nine vendors in particular rose to the top of the froth, exemplify this ecosystem in three areas: AI neoclouds, AI performance and support for the AI software ecosystem.

Here are my picks:

The burgeoning AI neocloud marketplace

Implementing AI strategies at scale requires a massive investment in infrastructure – graphics processing units or GPUs with all their data center trimmings.

Several companies are building out AI-centered hyperscaler (or hyperscaler-to-be) cloud data center offerings – aka “neoclouds” – that provide such infrastructure in the elastic, pay-as-you-go model of today’s more general purpose clouds.

The three neoclouds I interviewed are CoreWeave Inc., Nscale Global Holdings Ltd. and NeevAI SuperCloud Private Limited dba NeevCloud (with a shout-out to Nebius B.V., which I covered at a conference last year).

Of these three, CoreWeave has the most data centers (38 at last count). Its differentiator is building out racks of brand new GPU models to meet the needs of early adopter customers, even though fresh-off-the-fab GPUs often suffer from immature firmware and other growing pains. CoreWeave has built its stack to deliver resilient capacity despite these limitations.

Nscale is ramping up its data centers for a European market interested in data sovereignty and other Europe-specific compliance requirements.

Nscale’s differentiated approach to building out its neocloud is to arrange contracts for electric power ahead of demand to ensure availability and to manage future power costs. In other words, Nscale is placing its bets on power being the primary bottleneck to building out neocloud infrastructure.

While Nscale focuses on Europe, NeevCloud’s focus is on India. The company hopes to help its global customers build out AI solutions for the massive Indian market. It brings expertise navigating India’s particular regulatory environment while leveraging substantial government tax credits. It also sources hardware, bandwidth, power, and talent locally to keep costs down.

Souping up AI performance

Given AI’s substantial processing and data requirements, a central focus of today’s exploding AI ecosystem is resolving bottlenecks to increase performance.

Three vendors at RAISE stood out for their differentiated strategies for increasing AI performance. Ori Industries 1 Ltd offers a high-performance AI control plane that abstracts AI compute resources (in the cloud, on premises, or air gapped) and supports whatever large language models or LLMs and custom models its customers wish to run.

This control plane provides GPU virtualization, multitenancy, autoscaling and isolation, in addition to cloud native orchestration and machine learning operations or MLOps for strict governance and compliance requirements. As a result, customers can serve production models with the performance, control, and security they require.

Squeezing every nanosecond of performance from the underlying storage infrastructure also requires innovative approaches to managing and delivering storage. WekaIO Inc., which does business as Weka, is leveraging its expertise in the high performance software-defined storage market to offer NeuralMesh, a storage system it purpose-built to accelerate AI at scale.

Weka leverages the dedicated compute network that connects GPUs to each other to deliver wire speed performance – essentially offering speeds comparable to accessing memory via nonvolatile memory express or nVME storage technology.

The AI performance buck stops, however, on the chips themselves – the processors that make the magic happen. While GPUs garner the spotlight, one vendor that is innovating at the chip level is SambaNova Inc.

SambaNova offers high-performance AI with its ‘chips to models’ strategy. Unlike the other vendors on this list, SambaNova offers its own high-performance, low power reconfigurable dataflow unit or RDU chips as an alternative to GPUs.

SambaNova designed its RDU chips for the specific processing requirements of neural nets, eliminating the processing bottlenecks that limit GPU performance.

The company puts these chips into racks with its own operating system and networking software. These racks support all open source LLMs via OpenAI-compatible APIs. The result: vertically integrated AI stacks that support high end AI initiatives.

At RAISE, SambaNova also announced SambaManaged, its new inference-optimized data center offering that it designed for rapid deployment.

AI ecosystem supporting players

Three additional vendors at RAISE brought home this AI ecosystem story, as they each solve a particular set of problems within the broader landscape of AI solutions.

Dealing with the inefficiencies of LLMs is one such problem. The biggest problem with LLMs is the fact that they are large – and thus consume massive quantities of processor, network, and power resources, quickly running up costs.

Small language models or SLMs can reduce this overhead, but how does an organization know when SLMs can replace the more general LLMs?

Datawizz Inc. solves this problem specifying and tuning SLMs that it builds to meet the particular customer need. The Datawizz platform sorts prompts into use cases and trains SLMs accordingly.

Each SLM does one thing well. The platform then acts as a router transparently routing requests from the end-user or the developer building AI-based solutions to the appropriate model. The SLMs that Datawizz trains can handle the most common prompts, while the platform routes unusual requests to an LLM.

This approach works well with use cases like insurance claims processing, where the AI model sees the same claims forms with similar data over and over again.

Testing LLMs – ensuring they deliver high-quality results that meet the business requirements – is another problem for AI initiatives.

LLM behavior is non-deterministic: the same prompt and other initial conditions that provide the context for the interaction may yield different results on different occasions. As a result, traditional software testing approaches don’t work well on LLMs.

RagMetrics Inc addresses this problem of LLM testing via automated evaluation loops that leverage synthetic data as well as specialized LLMs that can judge the behavior of the LLMs being tested.

RagMetrics can also help customers select the best language model for their purposes by enabling comparisons based on answer quality, latency, cost, and many other metrics.

Finally, no AI story today would be complete without a nod to agentic AI. AI agents are adaptable, autonomous routines that leverage AI to complete tasks. Under the covers, this adaptability requires agents to generate the code necessary to complete their tasks – what Daytona Platform Inc. calls ‘glue code.’

The company offers an elastic infrastructure for agents to run this glue code, essentially providing a ‘virtual computer’ that agents can interact with to accomplish their goals, analogous to a real computer that a human would use.

Agents may need to call APIs, fetch data, view files, interact with other agents, or perform many other sub-tasks. Daytona provides the secure infrastructure for executing these sub-tasks, as well as a sandbox and tools for agents in development.

The Intellyx take

In my World Summit AI article last October, I pointed out that business use case-focused AI offerings were a poor excuse for an AI conference. After all, each industry has its own conferences, so solutions within those industries would find a more receptive audience at one of those than at a general AI conference.

The buzz at RAISE Summit proved my point, as industry-specific AI solutions were in short supply. Instead, the focus of most vendors at the conference was how they supported the AI ecosystem, from chips to data centers to infrastructure software to model performance.

I attend industry-specific conferences as well, from last month’s telco-focused TM Form DTW Ignite to the media and entertainment-centric IBC coming up in the fall. You can rest assured there will be plenty of industry-focused AI stories for me to tell at such events. Stay tuned!

Jason Bloomberg is founder and managing director of Intellyx BV, which advises business leaders and technology vendors on their digital transformation strategies. He wrote this article for News. Nebius is an Intellyx customer. None of the other vendors mentioned in this article is an Intellyx customer. No AI was used to write this article.

Photo: SambaNova/X