Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

• Wi-Charge, a startup using light beams to charge devices through the air, has shut down — learn more below.

• ARR Doesn't Mean What It Used To

• OpenAI just unveiled its first AI chip. It beat Nvidia. — learn more below

This Week In Startups

🔗 Resources

ARR Doesn't Mean What It Used To

How to Build a High-Growth Operating System

Execution Is The Moat

You are not a model. Don’t price per token.

📰 News

Chinese lab Z.ai says it’s behind the Ox Alpha model that wowed Silicon Valley

Apple debuts its ‘most powerful chip ever’

There’s a new Twitter in town

Nvidia closes in on Hugging Face acquisition

💸 Fundraising

Higgsfield, an AI image and video creation platform, raises a $400M Series B, Higgsfield raises $400M Series B

Groq, an AI inference cloud provider, raises a $350M Series A, Groq closes $350M Series A

Wispr, an AI voice dictation and interaction platform, raises a $280M Series B, Wispr’s Series B

Rillet, an AI-native accounting and ERP platform, raises a $100M Series C, Rillet raises $100M Series C

Fail(St)ory

Charging Through the Air

In 2015, a Samsung Galaxy S4 spun on a platform at Mobile World Congress while an invisible beam charged it from across the room.

That was Wi-Charge’s pitch in one neat demo: electricity could travel through the air and find a device without a cable trailing across the floor. In July 2026, the company filed for insolvency.

A co-founder later described the trap. Device makers wanted rooms full of transmitters before adding receivers to their products. Buyers wanted compatible devices before installing transmitters. Each side waited for the other, which is a difficult way to sell electricity.

What Was Wi-Charge:

Picture a hotel corridor full of smart locks. Someone has to notice the dying batteries, visit each door and replace them. Wi-Charge wanted one transmitter to keep several locks charged from across the hall.

Its AirCord system sent focused infrared light from a powered transmitter to a receiver built into the lock, display or charger. The receiver turned that light back into electricity. People could not see the beam, although walking through it stopped the power until the path cleared.

Devices carried a small rechargeable battery to survive those interruptions. The power also needed a direct view of the receiver. It could cross a room, but it could not turn a corner, pass through a pocket or negotiate with a badly placed cupboard.

The first working transmitter cost about $500,000 and required six PhDs to package, according to the co-founder. Years of engineering shrank later versions to roughly the size of a soft-drink can.

For stores, the appeal was movable digital signs without new wiring. Hotels and offices could avoid rounds of battery replacement for locks. Wi-Charge also demonstrated powered speakers, faucets, toothbrushes and a charging pad whose own power arrived wirelessly.

WIRED’s five-week bathroom test captured the experience nicely. A ceiling transmitter kept two electric toothbrushes charged from six feet away. Installing this wireless future required an electrician, an outlet in the loft and a hole in the ceiling.

It also used much more electricity than an ordinary toothbrush charger. Wi-Charge said the telemetry-heavy test unit was inefficient and a finished version would cut consumption sharply.

The company offered transmitters, receivers and integration kits to device makers, then built some finished products itself. These included store displays, an Alfred smart lock option and a $149 kit for Schlage Encode locks.

A co-founder claimed a footprint across 39 U.S. states, and the Schlage-compatible kit’s first batch reportedly sold out. Wi-Charge never disclosed that batch’s size, its revenue, margins or number of paid deployments.

The Numbers:

• 🧪 First transmitter: About $500K and six PhDs

• 📡 Published range: Up to 33 feet

• 💵 Consumer lock kit: $149

• 💰 Series C: $20M announced in May 2025

• 👥 Team: About 45 people in a 2024 estimate

• 🛑 Insolvency filing: July 14, 2026

Reasons for Failure: 

• Every room needed its own little network: Device makers told the co-founder they wanted a large installed base of transmitters before adding receivers. Property owners saw little value in transmitters while few compatible products existed. Wi-Charge had to persuade both sides to move together.

• The annoyance rarely earned a new system: Batteries and cables already powered the same locks, signs and toothbrushes. Replacing them was tedious, but the co-founder said customers rarely treated it as urgent. An invisible laser was much cooler than changing batteries; cool alone did not make the installation essential.

• Installation stayed physical: Each site needed powered transmitters, clear sightlines and receivers inside every target device. New product categories brought more integration work, from locks to bathroom fixtures and retail screens. The public record suggests growth kept depending on tailored deployments instead of one repeatable setup.

• Commercialization consumed a decade: Wi-Charge expected smart-home products in 2016 and a phone kit in 2017. It was still describing deployments as early in 2023, and its first direct consumer product arrived months before insolvency. The company announced a $20 million Series C in 2025, although the amount funded and its finances at closure remain unknown.

Why It Matters: 

• Deep-tech demos compress years of engineering into a few delightful seconds. The buyer pays for the whole package around them: installation, compatible devices, maintenance and a changed workflow. Test that package before scaling the hardware.

• Products that need infrastructure and compatible devices should measure useful density inside one site. Ten scattered pilots can look like traction while leaving the eleventh customer with the same empty-room problem.

Trend

OpenAI Goes Silicon

OpenAI says its need for compute is so large that it cannot imagine when it will have enough Jalapeño chips to sell one. For now, the chip is still being tested in its labs.

Jalapeño is OpenAI's first custom inference chip, built with Broadcom to serve ChatGPT, Codex, the API, and future agents. The strategy behind it is to use OpenAI's models and production workloads to improve the hardware, then use that hardware to make the products faster and cheaper to run.

At Hot Chips this week, the company published the chip's first measured performance results. Jalapeño ran three large open models and beat Nvidia's GB200 and GB300 on speed and energy efficiency.

Why it Matters

• Every ChatGPT response carries a compute bill. Custom silicon gives OpenAI another way to reduce that bill and another source of capacity.

• Latency is part of the product. A voice model that answers half a beat sooner feels better; a coding agent that can run efficiently for longer can take on larger jobs.

• OpenAI can now design around its own bottlenecks and product roadmap. The company says it will continue buying Nvidia hardware at scale. Jalapeño gives it more control over how quickly its infrastructure follows new models.

Jalapeño

An LLM request is several hardware problems disguised as one answer. Reading the prompt requires heavy computation. Producing each new token requires repeatedly pulling model data through memory. Spreading the work across chips adds communication delays.

Jalapeño handles those phases on the same system. It keeps the model's working memory local, changes which compute, memory, and networking resources are active, and powers down unused units. OpenAI designed the system around traffic from its own products, which shows where time and electricity are being spent.

Each Jalapeño package can hold 216 GiB of model data and move 15.4 TB of it every second. A full pod connects 2,048 chips, so you can do the math (it’s a lot). Each chip is rated at 700 watts, though OpenAI says the benchmarked workloads never pushed it beyond 550.

OpenAI then put Jalapeño against Nvidia's GB200 and GB300 using three large open models. Its chip completed 1.5 to 1.9 times more work per watt and returned full answers 1.7 to 3.6 times faster. In the tests designed to measure highly interactive use, Jalapeño was 2.1 to 4.1 times faster. On OpenAI's scorecard, Nvidia lost every matchup.

SemiAnalysis ran its public InferenceX benchmark with OpenAI engineers in OpenAI's lab and concluded that Jalapeño beat every Nvidia, AMD, and Google chip it had tested across several major open models.

The Trend

Jalapeño is the chip nobody can buy. OpenAI says it has no plans to sell it, and its internal demand is too large to leave much spare capacity anyway. Any advantage will show up inside ChatGPT, Codex, and the API.

That changes the economics of the AI race. If Jalapeño delivers more work per watt at scale, OpenAI can use the gain to lower prices, increase capacity, run agents for longer, or keep more margin on the same products. Competitors without comparable private infrastructure would still be paying a supplier for that layer.

A rival such as Anthropic could build an equally capable model and face worse serving economics. Matching OpenAI would then require better supplier terms, a more efficient model, or private silicon of its own. Pricing and scale would depend on the model and the infrastructure hidden underneath it.

OpenAI has also joined the chip industry from the buyer's side. Jalapeño is intended for its own data centers. As deployment grows, each chip can shift inference demand away from Nvidia and other suppliers. Those companies now face a customer capable of becoming its own supplier for part of an enormous inference load.

The scale remains unknown. OpenAI plans only limited Jalapeño capacity in 2026, will keep buying Nvidia hardware, and has published no cost per token, manufacturing yield, or production volume. Its 2027 rollout will supply the missing evidence.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico