HARDWARE

AMD’s Venice chips reshape the AI compute race

AMD's comeback story keeps getting stronger, and it couldn't come at a better time for an AI industry struggling to meet out-of-control demand for compute. 

On Thursday, AMD unveiled its next-generation "Venice" CPUs, officially dubbed the EPYC 9006 Series. It's a whole family of products aimed at matching the right CPU with the right workloads. These server CPUs are especially aimed at AI agents and high-performance computing (HPC).

The AI industry is paying attention to what Venice will mean for agents, which have exploded in popularity this year and deepened the compute crisis by massively increasing demand for infrastructure. Research has shown that agents can consume 1,000x more tokens than standard chatbot queries. 

And while large language models and chatbots rely heavily on GPUs, where Nvidia is the clear market leader, agentic workloads lean much more on CPUs for orchestration, feeding GPUs, and powering the enterprise services and tools that agents use to get work done. 

And in server CPUs, AMD rose to 46% market share of x86 server revenue and 33% of unit shipments in Q1 2026, according to Mercury Research. That's a continuation of its remarkable decade-long comeback story under CEO Lisa Su. Su, who became CEO in 2014, decided in 2017 that AMD needed to re-enter the server CPU market and make it a top priority. At that point, its market share had fallen to 0%. So approaching 50% nine years later speaks to how well the market has responded to the quality of its products. 

"I'm very happy to say Venice is in full production," AMD CEO Lisa Su announced on Thursday at the keynote of its Advancing AI event in San Francisco. "Customer demand is incredible. It's the strongest we've ever seen for a new EPYC generation. We're seeing every major server OEM and every major cloud provider on track to begin rolling out in the fourth quarter, as we start with the broadest EPYC launch we've ever had." 

The performance of Venice has even surprised AMD, which found that the high-end version of its chip offers 2.2x the performance per core than Nvidia's comparable Vera processor. That kind of power and efficiency will be welcomed by hyperscalers and neoscalers looking to boost performance and/or reduce the cost of their intelligence-per-watt. 

While AMD now offers Helios as a full rack-scale AI system to match Nvidia's industry-leading Grace Blackwell and Vera Rubin racks, the reality is that many server rooms and cloud providers go with best-of-breed options for AI by running AMD CPUs and Nvidia GPUs. That includes leading hyperscalers AWS, Microsoft Azure, and Oracle and neoscaler Lambda.

Make no mistake, Nvidia remains the dominant player in AI accelerators with 90% market share, because those are still mostly focused on GPUs. And Nvidia will likely remain the leader there for years to come. But AI workloads are transforming with the rise of agents, which will foundationally shift the compute and infrastructure needed to power AI. The voracious demand for compute shows no sign of slowing down, but the industry is becoming much more conscious about efficiency and cost. The pressure will continue to be on companies like AMD and Nvidia to deliver breakthroughs in performance while lowering the cost of intelligence-per-watt. The other thing to keep an eye on is how AMD's more open ecosystem play (ROCm) will continue to contrast with Nvidia's vertical integration (CUDA) to attract different players and partners, and what that part of the rivalry will mean for the shape of the AI ecosystem as it evolves. 

Disclaimer: Jason Hiner's travel to AMD's Advancing AI 2026 event was paid for by AMD. The Deep View's coverage is editorially independent from the companies we cover.

TOGETHER WITH IBM

What comes after tokenmaxxing?

AI adoption once centered on tokenmaxxing, with organizations driving more usage across development teams. But many are discovering that activity alone does not guarantee results.

Enter valuemaxxing, a shift toward measuring AI outcomes, including fast delivery, code quality, reduced rework and business impact.

Discover why maximizing value is becoming more important than maximizing usage.

Read the full article

PRODUCTS

Why the AI voice race is really about adoption

AI's two hottest labs are betting that voice is going to make AI better and easier to use for a lot more people. 

On Thursday, both Anthropic and OpenAI launched updates to their voice AI tools, allowing users more control over their models through speaking. Both updates allow AI systems to integrate more deeply into a user's context, enabling more complex task handling through free-flowing conversations. 

Here's the breakdown: 

• OpenAI announced that GPT-Live, its AI voice model that powers ChatGPT Voice, is now available in the ChatGPT desktop app, allowing it to work with agents via ChatGPT Work and Codex. This lets users direct agents to take actions across their computers and on the web using just their voice, including instructing agents to start and carry out multiple tasks. The update gives ChatGPT Voice access to plugins, files, app shots and computer use, giving it a better understanding of a user's context. 

• Anthropic's update brings Claude's voice mode to three different levels of models: Haiku for simple tasks, and Opus and Sonnet for harder problem-solving. Users can also now connect their voice capabilities to apps like Gmail and Slack, giving it more context for better task completion, and it's available in a wider array of languages. 

These updates are only the latest signal that voice is becoming one of AI's next growth areas as speech detection models shift the way that people work. OpenAI made this clear with its launch of Codex Micro, a miniature keyboard built for vibe coding that's controlled primarily via voice commands. 

And rumblings of this shift had already begun earlier this year at CES, where a large pool of startups showed off AI voice companions, from voice-powered earbuds to rings to hands-free, voice-controlled manufacturing use cases.

The name of the game in getting people to adopt any new tech is seamlessness. This is exactly why Anthropic, OpenAI and other startups are leaning into voice. It's a technology that can pick up and make sense of your streams of consciousness throughout the workday, and that could become incredibly valuable to almost any user. Getting as many people as possible to adopt AI into their day-to-day lives is critical for OpenAI and Anthropic to shepherd the AI transformation they have promised, as well as meet revenue goals ahead of two of the most high-profile IPOs in tech history. However, the seamlessness strategy only works if their voice AI is actually seamless. If not, it may have the opposite effect of what they intended, frustrating and turning off more users, as Apple has learned during its rocky history with Siri. 

TOGETHER WITH GRANOLA

The best AI notetaker is the one you barely notice

Most AI meeting tools are too visible: bots join your calls, notifications interrupt you, and afterwards you're left with a pile of transcripts.

Granola was built differently. It's an AI notepad that stays in the background so you can stay present in the meeting. You write down what matters, and Granola quietly captures the rest, turning the conversation into summaries, next steps, and context you can actually use.

You can even chat with your notes to write follow-up emails, prep for other meetings, or turn conversations into actions. Giving you, not your bot, a little boost.

Try Granola free for a month with code THE100OFF.

WORKFORCE

The best AI users don't outsource thinking

AI doesn't automatically translate into better work. But using it intentionally can.

A new study between KPMG and the University of Texas at Austin found that early-career professionals produced the strongest results when they treated AI as a thinking partner, rather than simply outsourcing work to it. Researchers found that workers who used AI to interrogate their work, challenge assumptions, and iterate on results outperformed AI on its own, revealing the limitations of passive AI usage. 

Researchers gave 523 early-career professionals access to an AI agent and asked them to complete role-specific tasks. They then evaluated participants based on both the quality of their work and how they interacted with the AI system, comparing their performance against the AI's standalone output.

• The researchers identified three distinct approaches to using AI. Half of the participants were classified as "AI amplifiers," who actively steered AI throughout the task. As a group, they produced work that outperformed the AI baseline.

• The other two groups saw weaker results. "AI delegators," who accepted AI's suggestions with little scrutiny, performed about as well as the AI working alone. "AI apprentices," who engaged critically with AI but often redirected it toward weaker answers, scored below the AI baseline.

• The findings suggest that foundational skills, defined in the study as critical thinking, domain knowledge, and AI literacy, remain essential to effective workplaces. But they aren't what separates the highest performers. 

"We weren't simply looking for people who knew how to use AI," Ashish Agarwal, professor at The University of Texas at Austin and co-author of the study, said in a release. "We wanted to understand what enables some individuals to consistently create value beyond what AI can produce on its own."

Workers who paired those skills with deliberate AI collaboration consistently produced the strongest results. It’s a takeaway that researchers suggest companies must consider if they’re looking to reap the benefits of AI.

"This is the most AI-native generation entering the workforce," Rahsaan Shears, who leads AI enterprise transformation at KPMG US, said in the release. "If fluency with the tools isn't what sets the top performers apart, that tells us something about our entire workforce."

Using AI with deliberate intention may be what separates quality work from slop. As companies increasingly mandate employees use AI to boost productivity, emerging research suggests the technology can backfire when used poorly. In some cases, AI is creating extra work as employees teach themselves how to use it and spend additional time correcting inaccurate or incomplete outputs. At the same time, some workers are becoming overreliant on AI to complete tasks, raising concerns about deskilling and the erosion of critical thinking and human judgment.  It's difficult to pinpoint the threshold that constitutes responsible AI use. But if companies insist employees incorporate AI into their work, they must be equally invested in helping them use it well. That means developing training that strengthens the human skills that make AI most effective.

LINKS

• Defense department signs 10-year deal with Oracle worth $7 billion

• AI hardware startup Etched raises $300 million at $10.3 billion valuation

• Yelp, ChatGPT partner to surface local, real-time reviews

• Cognition acquires Interaction, the makers of agent software Poke

• Microsoft nixes OpenAI's image models for its own in PowerPoint, Bing

• Samsung's foldable phones bet that AI can push the form factor

• Qwen-Image-3.0: New model from Alibaba, the third generation of its foundational image generation model. 

• Runway Media Router: The AI video startup launched a tool that automatically selects the best image, video, or audio generation model.

• ChatGPT Health: OpenAI's health tool is now available to all US users over 18. 

• FLUX 3: The latest model from Black Forest Labs, capable of generating images, video and audio, as well as use in robotic action.

• Gemini Spark: Google expands access to personal AI agent Gemini Spark.

• Postman: Principal Offensive Security Engineer

• Google DeepMind: Senior Security Engineer, Agentic Red Team

• OpenAI: Agent Post-Training Research

• Okta: Staff Product Security Engineer

GAMES

POLL RESULTS

Do you think your organization is prepared to handle AI-enabled cyberattacks?

Yes (15%)
Somewhat (22%)
No (53%)
Other (10%)