The rapid rise in the cost of using frontier artificial intelligence models is reshaping the enterprise AI market, creating an opening for infrastructure providers that help companies deploy cheaper open-source models while retaining greater control over their proprietary data.

One of the biggest beneficiaries is Fireworks, an Nvidia-backed AI cloud startup that said it has surpassed a $1 billion annualized revenue run rate, underscoring how demand is shifting beyond the largest AI model developers.

The San Mateo, California-based company announced Thursday that it has raised $1.5 billion in fresh funding at a $17.5 billion valuation, a fivefold increase in annualized revenue from a year ago.

“We’re seeing super-linear demand,” Fireworks co-founder and Chief Executive Officer Lin Qiao told CNBC. “This is a once-in-a-lifetime opportunity to have this kind of market.”

The milestone is notable because it highlights a growing segment of the AI ecosystem that is benefiting from enterprise concerns over both the cost and strategic implications of relying exclusively on proprietary models from companies such as OpenAI and Anthropic.

While OpenAI and Anthropic have each attracted valuations exceeding $800 billion this year, Fireworks’ rapid growth suggests that many businesses are increasingly looking beyond frontier AI labs in search of lower-cost and more customizable alternatives.

Fireworks operates in the inference cloud market, providing infrastructure that allows developers to deploy and optimize open-source AI models rather than relying solely on proprietary systems. The company competes with major cloud providers including Amazon and Google, as well as AI infrastructure startups such as Baseten and Together AI.

It has also expanded into GPU infrastructure for AI model training, putting it alongside specialized cloud providers such as CoreWeave, Lambda and Nebius.

Its growth reflects broader changes in enterprise AI spending.

Chief financial officers and technology executives are becoming more focused on controlling AI costs as model usage scales. Instead of paying premium prices for every AI workload, many companies are selectively deploying open-weight models for specialized applications where comparable performance can be achieved at substantially lower costs.

“Our cost compared with the equivalent-quality closed model is five to 10 times cheaper,” Qiao said.

The trend aligns with comments made this week by Microsoft Chief Executive Satya Nadella, who argued that companies should be able to use AI models without surrendering the institutional knowledge that gives them a competitive advantage. Nadella’s remarks echoed recent comments by Palantir CEO Alex Karp, who said enterprises increasingly want to “own the means of production” instead of depending entirely on external AI providers.

Fireworks has positioned itself squarely around that proposition. Rather than competing directly with OpenAI or Anthropic in developing frontier foundation models, it enables customers to build customized AI systems using their own proprietary data.

While OpenAI and Anthropic provide what Qiao described as “generalized intelligence,” Fireworks aims to deliver “specialized intelligence” by allowing enterprises to refine open models for specific industry use cases.

The platform hosts a wide range of open-source models, including offerings from Chinese developers DeepSeek, MiniMax and Z.ai, as well as open-weight models released by OpenAI. Customers can combine these models with proprietary enterprise data to improve performance for domain-specific tasks without exposing sensitive information to external model providers.

The company’s expanding partnerships also illustrate how the AI infrastructure market is becoming more collaborative rather than winner-take-all.

In March, Fireworks announced a partnership with Microsoft that integrates its services with Microsoft’s Foundry platform. The arrangement gives Microsoft customers access to Fireworks’ infrastructure while allowing Fireworks to leverage Microsoft’s enterprise distribution network.

“Through Microsoft we can get much bigger reach,” Qiao said.

The company relies on computing capacity from more than 20 infrastructure providers, including Microsoft, giving customers flexibility beyond the traditional hyperscale cloud model.

Its rapid expansion also challenges assumptions that Amazon, Microsoft, and Google will dominate every layer of enterprise AI infrastructure. Investors have increasingly rewarded companies serving specialized AI workloads. DigitalOcean shares have climbed 149% this year as AI-related demand accelerated, while GPU cloud provider CoreWeave, which raised $1.5 billion through its IPO last year, now commands a market capitalization of about $42 billion.

Fireworks’ operational scale has expanded dramatically alongside its financial growth. The company now processes approximately 40 trillion AI tokens every day, according to Qiao.

For comparison, Google disclosed in May that developers process roughly 19 billion tokens per minute through its AI models, equivalent to more than 27 trillion tokens daily. OpenAI said in March that its developer platform handled around 15 billion tokens per minute, or roughly 22 trillion tokens per day.

Although token counts are not directly comparable because pricing, workloads, and model architectures differ, the figures underline how quickly Fireworks has become one of the industry’s largest AI inference platforms. Its customer base has also diversified significantly. Last year, AI coding startup Cursor accounted for about half of Fireworks’ revenue. Today, Qiao said the business is substantially more diversified.

Cursor itself has reduced its dependence on OpenAI and Anthropic by developing its own proprietary Composer model. SpaceX agreed in June to acquire Cursor in a $60 billion stock transaction expected to close this quarter, further highlighting the strategic value of companies that control their own AI infrastructure rather than relying exclusively on external providers.

Other enterprise customers include Elastic, GitLab, and MongoDB.

Founded in 2022 by former Meta executive Lin Qiao and six co-founders, Fireworks currently employs about 200 people. The company plans to triple its workforce to around 600 employees by the end of 2026 as it expands its engineering organization, acquires additional GPUs and builds a dedicated enterprise sales force after years of relying primarily on self-service customer adoption.

The funding round was led by Atreides Management, Index Ventures and TCV, with participation from Nvidia, Evantic and Lightspeed Venture Partners.

Google on Thursday expanded the capabilities of its AI-powered search experience by allowing users to connect and interact with third-party applications directly within AI Mode, marking another step in its effort to transform search from an information tool into an AI assistant capable of completing real-world tasks.

The update, initially rolling out to users in the United States, introduces integrations with apps including Instacart, Canva and YouTube, enabling AI Mode to move beyond answering questions and helping users execute actions across multiple services without leaving the conversational interface.

The latest expansion is part of Google’s broader strategy of embedding artificial intelligence deeper into its search ecosystem as competition intensifies with OpenAI’s ChatGPT and Anthropic’s Claude, both of which have taken positions as AI assistants capable of interacting with external applications and services.

Rather than simply generating recommendations, AI Mode can now hand off tasks directly to partner apps. For example, users planning a barbecue can ask AI Mode to generate a grocery list and then connect their Instacart account to automatically add the ingredients to their shopping cart before completing the purchase through the Instacart app or website.

Similarly, users working on presentations or marketing materials can ask AI Mode to surface relevant Canva templates without separately searching the design platform, while those creating playlists can generate music recommendations and immediately save them to YouTube Music.

The integrations signal Google’s ambition to make AI Mode a central hub for everyday digital activities, reducing the number of separate searches and app switches required to complete common tasks such as shopping, designing content, and organizing entertainment.

The move also strengthens Google’s competitive position in the rapidly evolving AI assistant market, where major technology companies are racing to create platforms that not only provide information but also perform actions on behalf of users.

OpenAI has steadily expanded ChatGPT’s ability to connect with external services and productivity tools, while Anthropic’s Claude has added integrations designed to support enterprise workflows. Google’s latest update narrows that competitive gap by leveraging its own search platform alongside a growing ecosystem of third-party applications.

Thursday’s announcement builds on capabilities unveiled during Google I/O earlier this year, when the company introduced support for connecting third-party apps to Gemini, its flagship AI assistant. Those integrations already include services such as Canva, OpenTable, Spark and Instacart, allowing Gemini to complete tasks across multiple platforms.

The expansion of AI Mode suggests Google is aligning the capabilities of Search and Gemini, blurring the distinction between traditional web search and conversational AI. Google has steadily enhanced AI Mode since its launch in early 2025. Recent additions include the ability to check whether products are available at nearby retail stores, giving users real-time inventory information before they shop.

The company has also introduced a side-by-side browsing experience that allows users to explore websites while continuing conversations with AI Mode, preserving context for follow-up questions and product comparisons instead of requiring users to repeatedly start new searches.

Earlier this year, Google also introduced its “Personal Intelligence” capability, allowing AI Mode to access Gmail and Google Photos, with user permission, to deliver more personalized responses based on emails, travel plans, receipts, photos, and other personal information stored across Google’s ecosystem.

Together, the latest features underpin Google’s broader effort to reposition Search as an AI-native experience that combines conversational responses, personalized context and task execution. Instead of functioning solely as a gateway to websites, AI Mode is evolving into an orchestration layer capable of connecting multiple services, completing workflows, and keeping users within Google’s ecosystem for longer periods.

The strategy could also deepen user engagement with Google’s AI offerings while creating additional opportunities for developers and partner companies to integrate their services into conversational experiences. As more applications become compatible with AI Mode, analysts expect Google’s platform to serve as a unified interface for shopping, productivity, travel planning and entertainment, further reshaping how consumers interact with both search engines and third-party apps.

U.S. trade regulators have launched an investigation into Samsung Electronics’ memory chips and a range of artificial intelligence hardware sold by Google, Nvidia, Broadcom, and Super Micro Computer after California-based Netlist accused the companies of infringing its memory technology patents.

The investigation by the U.S. International Trade Commission (USITC) marks the latest escalation in Netlist’s years-long legal battle with Samsung and comes at a time when demand for high-performance memory chips is surging as hyperscale cloud providers and AI companies accelerate data center investments.

At the center of the dispute are dynamic random access memory (DRAM) technologies used in servers that power AI workloads. DRAM temporarily stores data for processors and has become one of the most critical components in AI infrastructure, particularly as advanced AI systems require enormous amounts of high-speed memory to train and run sophisticated models.

Netlist alleges that Samsung and its U.S. subsidiaries infringed patents covering data-processing technologies used in these memory products. Because Samsung supplies memory chips used across the AI hardware ecosystem, the complaint also names products imported by major technology companies, including Google, Nvidia, Broadcom, and Super Micro Computer, that incorporate the disputed components.

The USITC said an administrative law judge will conduct an evidentiary hearing before issuing an initial determination, which will then be subject to review by the full commission. The agency will establish a target completion date for the investigation within 45 days.

Netlist has asked the commission to issue one of its most powerful remedies: an exclusion order blocking imports of the allegedly infringing memory chips and AI products into the United States, as well as cease-and-desist orders preventing further sales.

If the commission ultimately grants those remedies, they would take effect immediately and become final after a 60-day presidential review period unless the U.S. Trade Representative vetoes them on public policy grounds, an action that is relatively rare.

The investigation highlights how intellectual property disputes are increasingly intersecting with the AI infrastructure boom. Demand for advanced memory chips has risen sharply as Microsoft, Amazon, Google, Meta, OpenAI and other AI developers continue investing hundreds of billions of dollars in AI data centers. That surge has fueled higher prices and record earnings for leading memory manufacturers, including Samsung, SK Hynix, and Micron.

High-bandwidth memory (HBM) and advanced DRAM have become strategic technologies because AI accelerators from Nvidia and other chipmakers rely heavily on large amounts of high-speed memory to maximize computing performance.

Although Nvidia, Google, Broadcom, and Super Micro are not accused of manufacturing the memory chips themselves, their products could become subject to import restrictions if the USITC determines that they incorporate infringing Samsung components. That raises the stakes considerably, as any import ban could potentially affect servers, AI systems and networking equipment sold into the U.S. market.

The case also exposes the growing legal risks facing the AI hardware supply chain, where a relatively small number of component suppliers serve virtually every major AI system builder.

The latest investigation builds on several courtroom victories Netlist has already secured against Samsung. In 2023, a Texas jury awarded Netlist $303 million in damages in a patent infringement case involving Samsung memory technology. That was followed in 2024 by another Texas jury verdict awarding Netlist an additional $118 million over related memory data-processing technology.

Rather than seeking only monetary damages through federal courts, Netlist is now pursuing trade remedies through the USITC, which has the authority to block imports into the United States. That approach is often viewed as powerful because it can create commercial pressure on technology companies even before damages are resolved.

The investigation follows memory technology’s growth as one of the most strategically important segments of the semiconductor industry. While AI processors from companies such as Nvidia often attract the most attention, industry analysts see advanced memory as an equally critical bottleneck for AI performance, making patent disputes in this segment potentially more consequential for the broader AI ecosystem than in previous semiconductor cycles.