GOVERNANCE

How the winds of tokenmaxxing shifted so quickly

As AI costs squeeze enterprises, the tokenmaxxers may be in for a reckoning. 

Tokenmaxxing, or the idea that heavy AI usage directly equates to enterprise value, has swept through Silicon Valley in 2026, with massive impacts. Gartner has predicted that AI spending will reach nearly $2.6 trillion this year, up 47% from the prior year. But those costs may be starting to hurt. 

Companies are questioning the spending sprees, and that includes some of the biggest names in tech. 

• In an interview with The New York Times this week, Microsoft CEO Satya Nadella said that being a tokenmaxxer is "addictive," but he's now pushing employees to use more efficient models, noting that they "don't use frontier models for non-frontier problems." This follows AI efficiency being center-stage in products announced at Microsoft Build last week. 

• Meanwhile, Uber CTO Neppalli Naga told The Information that the company burned through its entire Claude Code budget for the year by April. 

• And at the Nebius' Inflection forum this week, the company's CRO, Marc Boroditsky, said the industry needs to shift to "valuemaxxing," or juicing as much value as possible from your tokens, according to Alex Heath's Sources newsletter. 

• Nikita Shamgunov, VP of Databricks, said that customers are eyeing costs as “the wheels started to come off” of tokenmaxxing, Heath also reported. 

• Major AI firms may also be reading the tea leaves: According to The Wall Street Journal, both OpenAI and Anthropic are considering substantial price cuts for tokens as a means of winning customers from one another ahead of their record-breaking IPOs. 

So how did this problem emerge? According to Raj Ramanujam, VP of Global Alliances and Cloud at Dynatrace, enterprises jumped headfirst into AI out of excitement and fear of being left behind, often building their pipelines and workflows without thinking about the downstream costs. Now, he told The Deep View, "people are suddenly waking up to that in a very uncomfortable way." 

Rob May, CEO of Neurometric.ai, told The Deep View that this trend was born out of a simple need to measure the adoption and performance of AI in work settings. Many decided that measuring AI activity was a sufficient way to do so. However, while all of these tokens are measured in the same way, not all tokens are spent equally. For instance, a token that's used to write a to-do list may be counted equally to ones spent on complex tasks such as scientific research.

Additionally, the popularity of major model providers has led many to believe that other options, such as open-source models and small language models, aren't as viable. It's why May came up with an alternative: "Tokenminning." This is the idea that not every prompt and task requires ultra-powerful models, and that focusing on efficiency doesn't necessarily mean enterprises need to sacrifice on quality. 

"The big labs have done a great job of branding, and people believe they can do things that nobody else can do," he said. "I don't think that's true."

The money being poured into AI is nearly unprecedented, with both OpenAI and Anthropic approaching trillion-dollar valuations ahead of their public offerings, and xAI parent company SpaceX about to make its own debut at more than $1.7 trillion. Still, the question of whether the market is in a bubble continues to rumble under the surface. Now that enterprises have started to wake up to the reality of how much it costs to build AI into every part of their business, they may have to ask themselves whether some processes need the latest AI at all, saving the tech only for the most valuable applications. They may choose lower-cost SLMs and open-source models they can run locally for far lower costs and greater control. If that were to happen, the value of frontier models and frontier labs could fall back to earth.

TOGETHER WITH NEBIUS

Run open-source LLMs in real production

Capture live traffic, fine-tune and optimize, then deploy your own checkpoints to dedicated GPU endpoints.

Choose hardware, set scaling limits, and select region. Stable latency, predictable cost, clear data residency.

From LLM to production system, in one platform

BIG TECH

Apple fixed Siri, but that's not its biggest AI story

At Apple Park at the end of WWDC 2026, Sabrina Ortiz and I sat down for a special episode of The Deep View Conversations to break down Apple's biggest AI announcements and their impacts.

After two years of delays, promises, and mounting pressure, Apple finally delivered the Siri and Apple Intelligence experience we've been waiting for. But the bigger story may be how Apple is bringing AI to everyday users without forcing them into chatbots, new apps, or complicated workflows.

In this conversation, we unpack the most important announcements from WWDC, debate Apple's approach to AI, and explore where Apple's strategy differs from OpenAI, Google, Anthropic, and the rest of the industry.

Topics covered include:

• Why Apple's new Siri is a much bigger deal than it looks

• How 'Personal Context' is Apple's biggest AI advantage

• Why Apple is betting on features over chatbots

• The role of privacy, trust, and Private Cloud Compute

• Why Spatial Reframing is Apple's most innovative AI feature

• How Apple Intelligence works with Apple's Foundation Models and Google Gemini

• The AI upgrades coming to Photos, Safari, Messages, Writing Tools, and Mac

• What WWDC revealed about the future of AI on iPhone, Mac, Vision Pro, and beyond

• The biggest wins, misses, and unanswered questions from Apple's AI roadmap

If you're trying to understand where AI is actually headed, beyond the hype cycle, this is an episode you won't want to miss. Whether you're an Apple user, an AI enthusiast, a developer, or simply curious about how AI will show up in everyday life, this conversation offers one of the clearest perspectives on Apple's next chapter.

📺 Watch on YouTube
🎧 Listen in your favorite podcast player

Subscribe to Deep View Conversations for interviews with the leaders shaping the future of AI, business, and technology: tdv.transistor.fm 

TOGETHER WITH CODER

Self-Hosted Coding Agents. No Model Lock-In.

Coder Agents is a new way to run AI workflows, for teams that need visibility and control over where source code lives and where agents execute. It runs on self-hosted infrastructure, so platform teams get centralized governance and auditability without relying on hosted SaaS agent platforms.

Developers can delegate coding tasks, debugging, and research to persistent agents that continue running remotely across devices and sessions, even after the laptop closes. Coder Agents also supports any LLM provider, making it easy to switch between Claude, GPT, Gemini, or open-weight models without rebuilding workflows.

Get started with Coder Agents today and run AI workflows on infrastructure you control.

GOVERNANCE

Enterprise AI gaps are fueling 'Shadow AI'

As enterprises navigate their AI strategies, employees are using their own AI tools under the table. 

A report by Wakefield Research and PagerDuty published Thursday found that, of 1,250 office professionals surveyed, two-thirds reported using AI tools at work despite lacking explicit permission from their organization to do so. This phenomenon, known as shadow AI, has plagued workplaces since OpenAI's initial launch of ChatGPT, and can create significant data security vulnerabilities. 

According to the research, 88% of those surveyed have shared work-related information with public chatbots. Around 43% shared emails or other correspondence, 40% have shared meeting notes and summaries, 34%  have shared customer data, and 31% have shared financial information or confidential documents. 

Of those who have used shadow AI at work, 53% have received feedback to discontinue use, and 48% faced formal disciplinary action as a result. 

The survey points to several reasons for this undercover AI use: 

• Around 39% said they would prefer to use shadow AI rather than being told they can't use it at all. Meanwhile, 33% are avoiding scrutiny from management or leadership, and 30% are hiding it to avoid restrictive company policies and judgment from colleagues. 

• While 86% of organizations surveyed have AI policies in place, 81% believe that their leadership operates under a different set of AI rules than lower-level employees, perceiving a double standard. 

• Additionally, 72% of those surveyed are confident they understand AI better than their own tech teams do.

These driving factors, coupled with the fact that employees are wantonly putting information into public models, create a "massive enterprise liability," Tim Armandpour, CTO of PagerDuty, said in a statement. 

"We know the demand for AI is there because we see it in our own platform," Armandpour said. "The goal for any executive today should not be to slow down AI adoption, but to redirect that energy into proven platforms that offer governance and automation at scale."

AI has been improving at an exponential rate in recent months, as OpenAI, Anthropic and Google continue to leapfrog one another in capabilities. Though their models can do incredible things, they also present serious security risks when used in the wrong contexts. Additionally, because these models are changing so rapidly, it's becoming increasingly difficult for organizations and systems to keep up. It's why shadow AI — and employees' confidence that they fully understand a technology that even experts are struggling to grasp — can be so dangerous. Without more effective educational programs and a clearer picture of the consequences of using these models, employees are exposing their organizations to significant risk. 

LINKS

• Google commits $50 million for trade work training amid AI pressures  

• Amazon says its data centers used 2.5 billion gallons of water in 2025

• OpenAI to acquire Ona, an startup that offers cloud services for agents

• Agent platform Genspark raised $100 million at $2.6 billion valuation

• Deezer: Launches an AI music detection tool 

• DoorDash: Launches Ask DoorDash, an in-app chatbot 

• Perplexity Computer: Claude Fable 5 is available as an orchestrator model  

• Copilot Notebooks: Now rolling out to all Microsoft 365 Education users who use Copilot Chat, with any Academic license

• Salesforce: Applied AI Coach

• RAND: Senior Research Lead - AI Security Portfolio

• Applied Intuition: Research Intern - World-Action Foundation Model, Robotics

• Physical Intelligence: Applied Researcher

GAMES

A QUICK POLL BEFORE YOU GO