API of the tiger

Good morning,

Important news: the winner of the World’s Ugliest Dog contest was just crowned. Organizers say the contest is about showing beauty in imperfection. Beauty is in the eye of the leash holder.

In today’s edition:

• A new vendor category climbs the ranks
• The closed-door housing market
• The VC effect in AI adoption

The new middlemen of the AI economy rise up

One vendor category you’ll see a lot in this month’s list of top software vendors is “Model Serving and Inference.” While it doesn’t roll off the tongue, it’s trending hard.

You can think of these companies as the cloud hosting of the AI world. These vendors take trained AI models (usually, dozens if not hundreds of them) and run them on their own infrastructure for others. This is an especially useful service for companies that want to run several “open-weight” models like those from DeepSeek, Alibaba’s Qwen, or Meta’s Llama.

The reason these vendors are trending: companies want to be able to use AI more efficiently, which often means getting the best result at the lowest cost. And that often requires variety, or having more to choose from than just OpenAI and Anthropic.

• Why a go-between is useful: If a company wanted to run a full-sized model like DeepSeek on-premises, it would require data center hardware like GPUs (not to mention the round-the-clock maintenance that comes with that).
• So they outsource this inference to vendors like OpenRouter, Fireworks AI, and Baseten, which are among the fastest growing names on Ramp this month.

Businesses are increasingly using these platforms to manage their AI spend across models (think: automatically routing to the most efficient model for different tasks). They’re also using these services to be able to access cheaper open-weight models like China’s DeepSeek.

• Running inference: Stripe said it was acquiring OpenRouter (reportedly for $7.5B), which bills itself as “the unified interface for every model.” And Ramp recently launched its own AI model routing service.

Snowball effect… Companies are increasingly integrating AI features into their customer-facing products, but they can’t simply build these on top of a ChatGPT Plus subscription. Instead, they would use a model provider’s API, which is usage-billed. Ditto for heavy internal workflows like agentic research and coding. All these API-pinging use cases are billed by usage, and the more of them there are, the more opportunity there is for the vendors who act as the model-routing, inference-running middleman.

The bottom line:

Where there’s a bill, there’s a way. First there was the “tokenmaxxing” phase where companies scrambled to get up to speed with AI. Then there was the tokenmaxxing comedown where companies stressed over how much they were spending on AI and how little they could control it. Now, we’ve entered the solutions phase where companies are starting to look for cost efficiencies and visibility — while other companies are cropping up to address those very issues.

The closed-door housing market

This housing market is less about open houses and more about closed doors.

Lofty mortgage rates, high prices, low supply, and the “golden handcuffs” of pandemic-era rates mean that many are staying put — whether they want to or not.

Home Depot CFO Richard McPhail said the company is operating in “frozen housing market conditions.” While the retailer beat earnings expectations last week, McPhail said customers haven’t yet returned to big renovation projects.

Rival Lowe’s posted mixed results, citing pressure in home improvement spending. Both retailers are dealing with low home turnover, higher rates, and customers hesitant to splurge on major projects like a kitchen reno.

Homebuyers often drive renovation spend, and there aren’t a lot of them: a new report from Redfin estimates that the number of U.S. homebuyers has dropped to a record low, with sellers outnumbering buyers by 51% in July.

Pending home sales and contract signings dropped to the lowest level since January, with contract signings for previously owned homes matching the second-worst level in 25 years.

Homebuilding big shot Toll Brothers just reported that its sales fell year-over-year as it delivered fewer homes (though at higher prices).

With new home construction dropping, it could take a while for the closed-door market to open up.

Data Byte: the VC effect

Venture capital-backed businesses spend over 10x more on AI per employee than PE-backed firms. Venture-backed companies also lead in AI adoption (with an 80% adoption rate) as they influence portfolio companies to use the latest tools.

Signals Shortlist

• Trump admin unveils anti-Iran global sanctions plan, signals China not exempt (CNBC)
• Amazon hikes hardware prices by 60%, blaming memory shortage (TechCrunch)
• Moderna, Merck breakthrough could usher in wave of cancer vaccines (CNN via Yahoo Finance)
• Nvidia’s earnings tomorrow will test resurgent AI trade (Yahoo Finance)
• Shein targets $27B valuation in long-awaited Hong Kong IPO (Bloomberg)

🗓️ Leading Events:

Tuesday, August 25: Earnings expected from Dick’s Sporting Goods, Intuit, Zoom, and Semtech

Wednesday, August 26: GDP (2nd estimate). Earnings expected from Nvidia, Salesforce, CrowdStrike, Okta, and HP

Thursday, August 27: Earnings expected from Marvell, Workday, SentinelOne, Affirm, Best Buy, Dollar General, Dollar Tree, and Ulta

Friday, August 28: Fed Chair Warsh speech at Jackson Hole. University of Michigan consumer sentiment