RESEARCH

Running AI agents will cost 5x more by 2028

Everyone knows AI costs are rising. What fewer expected is that efficiency gains themselves may be driving the bill higher.

As improved efficiency lets research labs develop and deploy more powerful, more expensive models, and as users find increasingly sophisticated applications for them, like agentic workflows, token consumption keeps climbing, according to research firm Gartner. That combination is driving up overall inference costs, so much so that Gartner predicts AI inference costs per agentic workflow will increase more than fivefold through 2028.

This highlights an inference paradox, the subject of the report: even as unit economics improve, overall AI costs continue to rise without a clear or predictable path to matching value. The trap is that, while AI agents consume many more tokens, companies are encouraged to invest in these advanced assistants rather than AI chatbots because these agents are more likely to deliver returns and make step changes in the organizations. 

Even though those returns aren't predictable or guaranteed at the moment, and need to be higher than ever to justify the cost, the promise is enough to keep enterprises investing in AI solutions. While it would be most beneficial for companies to adopt a take-it-slow approach, outside pressures are unlikely to allow that, according to Scott Bickley, Advisory Fellow at Info-Tech Research Group. 

"The current environment has created a top-down fervor, in fact a mandate, for virtually all enterprises to aggressively adopt AI en masse," Bickley told The Deep View. "This blind foray into the AI abyss often lacks the in-depth understanding of the total cost of ownership, can ignore the culture of technology adoption within a given enterprise, and makes it difficult for one to advocate for anything but an 'innovator/early adopter' position." 

This demand is so insatiable that it has become the primary focus and biggest revenue generator for many leading AI labs, including OpenAI and Anthropic. As a result, these labs are also scrambling to find enough compute to meet demand. For instance, on Monday, news broke that Nvidia would provide up to $105 billion in financing for OpenAI's data center in Ohio. Nvidia highlighted in its blog post that AI factories are the "defining infrastructure" of the AI era, that "compute is revenue," and, as a result, it sees its responsibility to help secure those resources.

Typically, in any field, increases in innovation are regarded as positive. That's actually the beauty of technology: further developments lead to discoveries that would never have been possible without the previous advancements, creating a virtuous cycle. Yet with AI, it's a bit of a different story. Since AI was widely adopted, compute constraints have emerged, and they have only been exacerbated by growth in research on the development side. Compute and inference costs haven't kept up because both are based on finite, real resources. So, in a rare case for innovation, it may actually be wiser and more beneficial to pause, an idea supported by some of the biggest companies and experts in AI.

TOGETHER WITH VOICES

Find the Perfect Voice for Your Brand's AI

Training data quality determines model quality.

Voices delivers custom or pre-built voice datasets—directed by real professional talent, not scraped audio—across any language, accent, or emotional range. Choose from 800+ hours of character voice performances, 1,723 hours of QA'd conversational data across 56 emotional states, or 1,000 hours of expressive, multi-language data with full emotion tagging.

Every dataset ships with clear licensing, signed consent, and usage rights baked in, so your legal team sleeps easy and your model trains on data it can actually use. 7 of the world's top 10 AI labs already do.

Request free samples and hear the difference professionally directed voice data makes.

MARKETS

AI's coding gold rush is getting crowded

AI coding startup valuations are skyrocketing. How much higher can they go? 

Last week, three AI coding startups raked in massive funding rounds at multi-billion dollar valuations, a sign that investors are confident in the value and returns that AI-powered coding can bring to enterprises. 

Though much of the AI coding hype has thus far been centered around Anthropic's Claude Code, OpenAI's Codex, and now SpaceXAI-owned Cursor, other startups are emerging as competitors: 

• Code Rabbit, an AI code review platform focused on security and governance, raised a $143 million Series C at a $1.5 billion valuation, which it will use for international expansion, research and product development.

• Lovable, a European vibe coding platform, raised a $400 million Series C funding round at an eye-popping valuation of $13.3 billion. The round roughly doubles the company's previous valuation of $6.6 billion from its December funding round, and follows Lovable hitting  $500 million in annualized run rate revenue, according to TechCrunch.

• And Cognition, which operates Devin, an "autonomous software engineer," is in talks to raise $1 billion at a valuation of $40 billion, according to Bloomberg. The report follows the company raising $1 billion at a $26 billion valuation earlier this year as it reportedly approaches an annualized revenue run rate of $1 billion.  

Of all of the enterprise use cases for AI, coding has been the clear winner thus far. Big tech firms are seeing massive adoption. Sundar Pichai, CEO of Google, said in April that 75% of the company's code is now written by AI. Snap CEO Evan Spiegel, meanwhile, said that 65% of its code is AI-generated, and Airbnb CEO Brian Chesky said that 60% of the code created by the company's engineers is now written by AI.

Right now, many of the companies that are buying into AI coding are in an experimentation phase. At conferences, when I ask people about which coding platforms they use, they often say something along the lines of "all of them." However, that experimentation likely won't last forever. Think of it like streaming services: While many people used to subscribe to Netflix, Disney Plus, HBO Max, Paramount Plus, Peacock, and whatever else when they first launched, most eventually pared it down to two to three favorites. The same phenomenon could happen with coding tools: Once developers decide what the Netflix or HBO Max of AI coding tools are, they might stop paying for the rest. So while valuations are sky-high now, don't be surprised if they reach a tipping point. 

IN PARTNERSHIP WITH LAMBDA

Cut your AI training costs by 25% or more

Most large-scale AI training runs use less than half the computing power they're paying for. Lambda's team found the root causes and built a reproducible framework that boosted efficiency by over 25%, without changing the model itself.

Lambda’s whitepaper shows you how to address:

• Memory inefficiencies silently inflating your costs

• Training configurations that aren't making full use of your hardware

• Bottlenecks that slow down GPU communication

Get the whitepaper.

PRODUCTS

Why you might want an AI phone with a robot arm

For almost two decades, smartphones have grown smarter and more powerful, yet somehow ended up all looking the same: a glassy brick in your pocket. But things are starting to get fun.

I had a chance to spend time with Honor's new Robot Phone ahead of its launch. At first glance, it looks like an ordinary slab smartphone. But in its full glory, a robot arm on a gimbal emerges from the camera socket. The gains for content creation are clear. The robot head, equipped with a 200MP main camera, can track your movements for photos and videos, keeping your shot stable while you are still in frame. However, the most interesting applications of this little gimbal relate to AI.

During the briefing, Honor described it as an "AI companion." The claim is supported by the adorable (or dystopian, depending on your preferences) interactions that YOYO, the AI assistant, has with users, such as waking up to its name with a little dance. It also has expressive movements, such as nodding and shaking, accompanied by over 100 on-screen facial expressions.

While I do think that having such a cute interaction with an AI assistant may compel people to chat with AI more—and, admittedly, I quite enjoyed chatting with it myself—the real value lies in using AI tracking on the gimbal for multimodal AI prompts.

I've said this for years (even on national television): one of my favorite AI use cases is cooking. Since I like to cook off the cuff, showing an AI assistant my ingredients helps me get precise measurements. My hands are usually dirty, so I leave Gemini Live on and just move into frame when I need to ask something. But what if my phone's AI assistant could just follow me around the kitchen instead?

With the Robot Phone specifically, it's not as simple as that, as the phone would still have to be propped up on Honor's included kickstand, but it's a step in the right direction. It's the same reason I'm such a fan of Samsung's Galaxy Z foldables for AI. They make hands-free assistance easy. With the Galaxy Z Fold 8, which I've been testing for three weeks, you can set it up in tent mode so the outer screen faces you and the selfie camera stays exposed, letting you chat with AI hands-free while it still sees you.

To be clear, do I think most people need an Honor Robot Phone at the moment, or that phones with robot arms like it will become the new norm? Not at all. Rather, I give Honor credit for trying such a novel form factor for a smartphone.

The Honor Robot phone is only available in China, and it's clearly a 1.0 product. However, Apple is rumored to finally be entering the foldable market by releasing a device similar to the Galaxy Z Fold 8 in September, and the new Samsung foldable lineup is already available for purchase. So if you're interested in experimenting with new form factors and possibly elevating your AI game, you've got options.

Form factor is only half the story, of course. The AI itself needs to keep getting better in these phones. For instance, the Google Pixel 10 has been the best AI smartphone for the past year, despite being a slab, largely because Gemini is so seamlessly integrated into the phone and powers actually helpful features. Google also has its own foldable, the Pixel 10 Pro Fold, which benefits from the same hands-free advantages as other foldables I mentioned, though Google's version lags behind competitors like Samsung due to its heft and bulk. Its newest addition, the Pixel 11 Pro Fold, which just launched last week, is one to watch. If hardware upgrades make the foldable, hands-free form factor on par with Samsung's lineup, and combine that with the best AI built into a smartphone, it could truly claim the crown as the best AI phone. I will be testing it for the upcoming weeks and updating you as I go across my socials (Instagram, Twitter and Threads), as well as The Deep View.

LINKS

• Igor Babuschkin's River AI startup raises $1.1 billion to build user-owned AI

• OpenAI signs Ohio data center lease backed by Nvidia and SB Energy

• WSJ finds Big Tech's AI spending $3 trillion higher than documents show

• Greg Brockman calls OpenAI-Hugging Face incident a "watershed moment" 

• Wispr reaches $2 billion valuation as it eyes growth beyond dictation

• Wonder, Serve Robotics to launch robot delivery on Grubhub in three cities

• Codex: Users can enable a 1 million token context window in Codex for GPT-5.6 Sol 

• Z.ai: Unveils GLM-5.3, claiming coding and security gains

• Matic Cues: Voice and gesture control are now available for Matic's household robots.

• North Micro Vision: The newest model from Cohere, its smallest vision-language model yet, available open source under the Apache 2.0 license.

• The OpenAI Foundation: AI Deployment Engineer

• WorkOS: Applied AI Engineer

• Mistral: AI Scientist

• Variant: Applied Researcher

GAMES

A QUICK POLL BEFORE YOU GO