MARKETS

Why AI insiders are all in, but the public isn't

AI bubble talk is rearing its head again, but the context is very different from the conversations in late 2025.

In this episode of The Deep View Conversations, we unpack the common arguments about an AI bubble and explain why reality naturally falls somewhere in between the doomsayers and AI absolutists.

We look at AI's "Tinker Bell problem": the boom depends partly on people continuing to believe in AI's potential, even as public skepticism grows. Beneath that belief cushion, enterprise contracts drive most of AI labs' revenue, while strong hyperscaler earnings and compute shortages suggest durable demand is building.

We debunk a viral claim that a $200 Claude subscription costs Anthropic $8,000 to serve. We also look at enterprises' push for more control, efficiency and measurable ROI, including one company's claim that some engineers' token use costs 1.5 times their compensation.

Other topics include:
• Training, inference, API pricing and token economics
• Real value, snake oil and the hype cycle
• Why AI demand outruns compute supply
• Why the AI bubble may look more like bubble wrap
• Market rotation into energy and materials

If you're trying to separate durable AI demand from hype and understand where a real correction could begin, then this conversation offers a framework for thinking about what may pop, what may deflate and what may keep growing. Keep in mind that this is industry analysis and not investor advice.

📺 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.transitor.fm

TOGETHER WITH GRANOLA

Be The Person Who Remembers Everything

Ever notice those people who walk into meetings and somehow remember every conversation, decision, and next step from the last time? They’re not some kind of savant – they just use Granola. 

Granola’s AI notepad captures your meetings automatically. No bots, no awkward recording prompts, just a quietly powerful log of everything that matters. Ask Granola questions via chat and you’ll instantly get the answer you’re looking for, including who it came from, when it was solved, and what notes were taken on it. Best of all, the more you use it, the better Granola gets – over time, it builds a searchable memory of every discussion and decision made. 

If you’re ready to become the person in the room who remembers everything, it’s time to try Granola for free right here.

PRODUCTS

Chinese AI bets price can overcome trust

Chinese model firms have set their sights on a lucrative target: Enterprises. 

On Thursday, DeepSeek launched Harness v0.1, an open-source agent harness, in developer preview under the MIT license. Additionally, the firm launched DeepSeek-V4-Pro, the latest iteration of its flagship model built for agentic workloads. 

The release provides an open-source challenger to Anthropic's popular Claude Code and highlights that DeepSeek may be eyeing the developer audience in earnest. But it's not the only Chinese AI firm with an interest in the enterprise: 

• On Friday, Alibaba launched Qwen-3.8, a lightweight, 27-billion parameter model that outperforms its previous generations, particularly in "real-world coding and office workflows," the company said in its post on X. 

• Additionally, Z.ai released GLM-5.3, its latest flagship model, which it claims is its "most capable open-weights model for coding," with an over 50% improvement over GLM-5.2 on its in-house coding benchmark. Additionally, the model has "emergent" cyber capabilities, including improved vulnerability discovery.

These product releases come at a time when the conversation about Chinese open-source models is more fervent than ever. Largely, these models are more cost-efficient and customizable than competitors like OpenAI 's GPT-5.6 Sol and Anthropic's Fable 5, which are starting to hurt enterprise pocketbooks. 

But these releases signal that Chinese firms want a deeper cut of the enterprise market, potentially targeting enterprise adoption and coding as an entry point. And it makes sense why they might be zeroing in on business adoption: Frontier US AI labs continue to rake in money powered by enterprise contracts, with OpenAI's revenue run-rate projected to hit an eye-popping $40 billion before its IPO and Anthropic is eyeing a $2 trillion valuation ahead of its own public offering.

It makes sense that Chinese AI labs want a hold of the cash cow that is enterprise AI deployments. And they may be targeting this market at a particularly poignant time as many organizations grapple with the often dizzying cost of Claude Code and Codex bills. But these models come with a carrot and a stick. The carrot, of course, is cost. These open-source models are generally performant for many enterprise needs, and cost a fraction of their proprietary competitors. But the stick is trust. For one, major model labs, including OpenAI, Anthropic and Google, have accused each of these Chinese model providers of model distillation. Additionally, these models present their own security risks, including weaker guardrails and jailbreak risks. Because of this, enterprises must weigh the trade-offs and decide whether they feel comfortable embedding these models within their organizations. 

TOGETHER WITH JUMPCLOUD

Meet the Zombie Agents already inside your environment

An agent gets deployed. The project ends. The developer moves on. The agent keeps running, keeps accessing systems, and keeps accumulating permissions with no owner, no review, and no offboarding process.

Research from JumpCloud found that only 21% of organizations have governance controls in place for non-human identities. Zombie Agents are not a future risk.

Find out what securing every identity actually looks like →

RESEARCH

Data: AI agents hit an enterprise reality check

AI agents are pitched as digital coworkers that can work alongside teams. But new research suggests companies are struggling to make that vision a reality.

A new Deloitte survey of more than 500 U.S. business and IT leaders found a wide gap between companies’ ambitions for AI agents and their ability to actually use them at scale. Nearly three-quarters of leaders said they expect half of their businesses to be rebuilt or designed around AI agents within the next four years. However, even though 42% said their organizations have tested or deployed agents, just 15% have scaled multi-agent systems across the business.

The bottlenecks boil down to the architecture surrounding the agents: 

• Just 21% of leaders said their business processes are prepared for agentic AI, and only one in five said their organizations are ready to redesign processes for agents. 

• Instead, many companies are slapping agents onto existing ways of working in hopes of getting faster returns. 

• Leaders also pointed to inaccessible data, difficulty governing agents, and costly integrations as major barriers to scaling.

“Limited, layered-on approaches may create the sense of getting ahead with quick wins, but in reality, they may not be enough,” Laura Shact, Deloitte’s US Technology, Media and Telecommunications AI growth leader, said in the study. 

A separate study from Google and MIT Technology Review drills into another one of the biggest obstacles to using agents: data. More than half of the 300 global data and technology executives surveyed said legacy data systems are preventing them from scaling AI agents. On average, companies currently give AI access to just 45% of their enterprise data, which could limit its potential for efficiency gains. 

The challenge goes beyond simply opening up more databases. Companies said their data is scattered across disconnected systems, locked away in unstructured formats like emails and PDFs, unavailable in real time, or missing the business context agents need to understand what the information actually means. Those shortcomings become more consequential when agents are expected to make decisions and take actions instead of simply answering questions.

Different levels of data access can compromise the quality of agentic workflows. Among organizations that give AI access to more than 70% of their enterprise data, every respondent characterized their agents’ outputs and decisions as either mostly or consistently accurate. And of organizations giving AI access to 30% or less of their data, just 22% said they trust the accuracy and relevance of their agents’ decisions.

"The so-called 'modern' data stacks that are glued together with disjointed parts were not built for the agentic era," Andi Gutmans, Google Cloud’s vice president and general manager of Data Cloud, said in the study.

The bottlenecks point to a bigger problem: Most companies weren't built for autonomous software. Getting agents to work reliably at scale will require more than deploying better models. Companies will need to break down data silos, redesign workflows around what agents can actually do, put guardrails in place for when they act out of line, and train workers to use and oversee them. The technology may be moving fast, but the harder work will be getting the rest of the business ready for it.

LINKS

• Google makes visible AI content watermarks optional in Gemini

• Pony AI plans to deploy 2,000 robotaxis across Europe

• OpenAI reportedly on track for $40 billion ARR ahead of IPO

• Nevada grants Tesla robotaxi permit for just 10 vehicles, not 5,000

• Dario Amodei weighs in on open models, decentralization, and AI regulation

• Stripe moves to acquire OpenRouter for $7 billion

• Hugging Face publishes insights on the state of open models

• ChatGPT Computer History: OpenAI's flagship chatbot can now remember user activity across the apps and websites on a user's computer.

• Dreamina Seedance 2.5: The Chinese video generation model is now available in 1080p. 

• Palmyra X6: WRITER's new model cuts AI agent costs by 52%. 

• Claude Tag: Athropic's flagship model now has context to decide when to proactively collaborate in Slack.

• SafetyKit: Applied AI Engineer

• Figma: AI Applied Scientist

• Roblox: Technical Director - AI Roblox Studio

• Arcade.dev: Applied AI Engineer

GAMES

POLL RESULTS

Do you think AI costs are growing out of control at your organization?

Yes (23%)
Somewhat (30%)
No (40%)
Other (7%)