It’s widely acknowledged that AI data centers have been putting pressure on the grid. However, Silicon Valley has been somewhat shielded from these issues due to high land and electricity costs that have prompted hyperscaler initiatives to relocate. 

The tech aristocracy might soon experience the impacts of the power shortage, however. The Bay Area’s leisure destination, Lake Tahoe, has less than a year to secure a new energy provider.

By May 2027, the agreement between Liberty Utilities and NV Energy will conclude. NV Energy’s electricity will be redirected to other areas in Nevada, where the data center industry has been flourishing.

Both Liberty Utilities and NV Energy have indicated that this phase-out has been in the works for a long time, and NV Energy claims that data centers are not responsible. Yet, it is difficult to argue that they do not contribute. NV Energy alone has more than 22 gigawatts of load requests, which, as highlighted by a Bloomberg report, exceeds 40 times Lake Tahoe’s peak usage. 

If data centers were not a factor, it’s plausible to envision a scenario where Liberty Utilities and NV Energy extend their contract. However, with data center clients prepared to pay whatever is necessary for electricity, it was unavoidable that traditional consumers in Lake Tahoe would be left without power.

The timing is particularly unfortunate. The energy markets are currently challenging, strained by soaring demand and diminished supply further complicated by the previous administration’s confrontation with Iran.

Lake Tahoe’s situation is exacerbated by the reality that its power lines are more interconnected with Nevada’s grid than California’s. Consequently, the community must seek another electricity supplier from within NV Energy’s domain or beyond in the West. 

Given that NV Energy has already placed data centers above the mountain community in priority, it is probable that Lake Tahoe residents — along with second-home owners — will need to locate an alternative regional energy producer.

That won’t be a simple task. Just one state away, in Utah, a county commission has recently authorized a 40,000-acre data center project projected to consume up to 9 gigawatts of electricity upon completion. Currently, the entire state of Utah uses about 4 gigawatts. Demand on this scale is almost assured to inflate prices across the region.

The combination of these factors suggests that Lake Tahoe will likely incur higher electricity costs next year compared to today. Local residents will bear the brunt of this increase, but second homeowners in the area, many from Silicon Valley, may also feel the financial strain. 

The irony of the AI energy crisis is that those most affected have had minimal influence over the technology or its deployment. Lake Tahoe’s energy situation indicates that this dynamic is beginning to shift, though probably not significantly enough to effect real change.

Since July 2025, Tesla Robotaxis have experienced at least two crashes while being remotely driven by a teleoperator, as revealed by newly disclosed data submitted to the National Highway Traffic Safety Administration (NHTSA).

Both incidents took place in Austin, Texas, at low speeds. In each event, a safety monitor was in the driver’s seat and no passengers were present.

This new information was released just a few months after Tesla informed legislators that its remote operators could control one of the company’s vehicles as long as the speed remained below 10 miles per hour. “This capability allows Tesla to quickly maneuver a vehicle that might be in a tough situation, reducing the necessity to wait for first responders or Tesla field agents to manually retrieve the vehicle,” the company stated at that time.

Like other firms developing autonomous vehicle technologies, Tesla is obligated to report detailed information about any accidents to the NHTSA. However, unlike most others in that space, Tesla consistently redacted information about its crashes, claiming it was proprietary business information.

It remains uncertain what prompted the change, but Tesla shifted its stance this week, and the latest data released by the NHTSA now includes a detailed account of all 17 crashes Tesla has documented since the inception of its Robotaxi network last year.

In July 2025, shortly after launching the network in Austin, the company’s automated driving system (ADS) reportedly struggled to move forward while stopped on a road. The safety monitor sought assistance from Tesla’s remote help team, leading a teleoperator to “take control of the vehicle and gently accelerate, steering the Tesla ADS to the left side of the street.”

The teleoperator subsequently drove “up onto the curb and struck a metal fence.”

A similar incident occurred in January 2026. The Tesla ADS was proceeding straight on a roadway when the safety monitor “requested support for vehicle navigation.”

“The teleoperator assumed vehicle control when the ADS was halted and continued straight down the street. The Tesla vehicle collided with a temporary barricade set up for construction work at roughly 9MPH, damaging the front-left fender and tire,” according to the data submitted to the NHTSA.

In line with other companies in the autonomous vehicle sector like Waymo, a majority of the newly disclosed crashes involve Tesla Robotaxi vehicles being struck by other vehicles rather than causing accidents themselves.

However, at least two incidents involved a Tesla Robotaxi grazing its mirrors against other cars. One accident from September 2025 saw the Tesla ADS unable to avoid hitting a dog that dashed into the street. (Tesla stated that the dog managed to escape.)

In another September 2025 incident, a Tesla Robotaxi made an unprotected left turn into a parking lot and collided with a metal chain. (The NHTSA recently concluded an investigation into Tesla’s Full Self-Driving software, which occasionally crashes into parking lot barriers, chains, and gates. Waymo also issued a recall last year concerning a similar issue.)

While competitor robotaxi companies such as Waymo and Zoox have reported more accidents than Tesla, Elon Musk’s company operates at a significantly smaller scale. The details unveiled this week in the newly unredacted data may clarify why Tesla is cautiously expanding its early-stage autonomous ride-hailing service. Musk himself acknowledged last month that “ensuring total safety” is the primary limiting factor for Tesla’s expansion, stating that the company is proceeding with “great caution.”

On Friday, OpenAI introduced a fresh suite of personal finance tools in preview for ChatGPT Pro users in the U.S., allowing them to link their accounts and inquire about topics from spending insights to future financial strategies.

OpenAI has collaborated with the financial connection service Plaid to facilitate the account integrations. Users can link to more than 12,000 financial institutions, such as Schwab, Fidelity, Chase, Robinhood, American Express, and Capital One. After linking these accounts, users will be presented with a dashboard displaying their portfolio performance, expenditures, subscriptions, and future payments.

This new offering arrives just a month following OpenAI’s acquisition of the team associated with personal finance startup Hiro, which received backing from firms like Ribbit, General Catalyst, and Restive in April. OpenAI indicated that the financial expertise of the Hiro team contributed to the launch of this offering but did not clarify if the entire feature was developed by them.

OpenAI users can access the feature by clicking “Get started” under the “Finances” option in the sidebar, or by typing “@Finances, connect my accounts” during a ChatGPT conversation. After doing so, the chatbot will assist them in linking accounts via Plaid. The company mentioned plans to support Intuit soon, which would facilitate analysis such as the effects of a stock sale on taxes or the probabilities of credit card approval.

OpenAI reports that over 200 million users pose financial inquiries to ChatGPT each month. The company also highlighted that the new GPT-5.5 model excels at contextual reasoning, which is vital for addressing finance-related queries. The firm stated that it collaborated with finance professionals to establish a benchmark for enhancing the model’s responses to personal finance questions.

With the integration of the new financial tools, users can receive detailed responses to questions like “I feel like my spending has increased lately. Has anything changed?” or “Assist me in devising a plan to prepare for purchasing a home in my area within the next 5 years.”

Users can navigate to Settings > Apps > Finances to disconnect certain accounts if desired. Upon disconnections, the synced data will be purged from ChatGPT within 30 days. Additionally, users can review and delete financial memories from the Finances page.

Generalized chatbots are built to respond to any inquiries, prompting users to ask about sensitive data topics such as health, finance, and personal matters. AI firms are acknowledging this trend and developing specialized products for these domains. Both OpenAI and Anthropic have released tools related to health. Earlier this month, Perplexity introduced its own financial research tool driven by its Computer agent.

OpenAI stated that its personal finance tools will be accessible on ChatGPT on the web and iOS for Pro users. It emphasized that it intends to enhance the product based on feedback from these users before rolling it out to Plus users.

The PJM Interconnection, the largest electrical grid in the U.S., experienced prices that nearly doubled over the past year, as per a report released yesterday by Monitoring Analytics, an independent market overseer acting as a watchdog for the PJM grid. The cause? Data centers.

Wholesale electricity prices for one megawatt-hour surged to $136.53, increasing from $77.78 at the corresponding time last year. Crain’s Chicago Business was the first to cover the price surge. Monitoring Analytics pointed to data centers and PJM’s inability to adequately manage their increasing demand.

The market monitor did not hold back. “The price effects on customers have been extremely significant and cannot be reversed,” Monitoring Analytics stated. “The price effects will be even more pronounced in the short term unless the challenges related to data center load are resolved promptly.” 

PJM is susceptible to such critiques. In 2022, coinciding with the rise in data center construction, the grid operator halted applications for new generating sources, due to a lengthy backlog. It only recently began accepting new requests. Meanwhile, electricity consumption from data centers has surged dramatically. The PJM grid includes Northern Virginia, an area densely populated with data centers.

The price increase serves as a reminder of a more profound issue: The U.S. power grid was not equipped for the electricity demands of an AI-centered economy, and the disparity between what the grid can provide and the industry’s requirements is growing.

Monitoring Analytics asserted that without the heightened demand from data centers, “the capacity market would not have encountered the same stringent supply-demand conditions, nor the elevated prices observed.”

It further noted that “the current capacity supply in PJM is insufficient to satisfy the demand from large data center loads and will continue to be inadequate in the near future.”

Monitoring Analytics criticized PJM’s lack of transparency in decision-making and its delay in essential software upgrades. “These upgrades have faced multiple years of delays and lack a definitive expected implementation timeline,” the report indicated. 

The report follows a white paper published by PJM Interconnection, which explored the future of the grid it oversees. The white paper proposed three possible directions, but none attracted the interest of one of the region’s largest utilities, AEP, which has threatened to exit the PJM grid entirely.

Monitoring Analytics expressed similar dissatisfaction with PJM’s white paper. The group remarked that PJM was utilizing the crisis “as a pretext” for altering the operation of its power market. “The fundamental aspects of the PJM market design remain sound,” it stated, implying that the grid operator had mishandled its reaction to rising demand. The solution, it asserted, “begins with acknowledging that the root of the current problems is data center load.” In other words, it’s the data centers, plain and simple.

The productivity software creator Notion is entering the era of agentic capabilities.

During a live-streamed product launch on Wednesday, the organization, best recognized for its collaborative note-taking application, unveiled a new developer platform that enhances the functionalities of its tailored AI agents, integrates external agents, and enables teams to create automated multi-step workflows that can extract data from any database.

By establishing an orchestration layer — a framework that synchronizes AI tasks across various tools and data sources — Notion aims to redefine itself from merely a note-taking service with AI capabilities into a central hub for collaboration among individuals and agents across different tools and databases.

In February, Notion initially launched its Custom Agents — AI team members responsible for handling repetitive tasks such as responding to FAQs, compiling status reports, and automating workflows. Since their introduction, Notion reports that more than one million agents have been created by its users.

Nonetheless, these agents faced certain restrictions. They were unable to connect to outside data or apply custom logic. Additionally, external agents utilized by companies lacked a means to interact with the Notion workspace, forcing teams to navigate these challenges through third-party automation platforms or custom scripts operated on their own infrastructure.

“Historically, it’s accurate that Notion hasn’t been the most developer-centric platform,” remarked Ivan Zhao, co-founder and CEO of Notion, during the livestream. “But that’s changing.”

Now, Notion will empower teams to implement their own custom code. With the introduction of its new Workers, a cloud-based infrastructure for executing custom code, clients can craft their logic and deploy it within a secure sandbox (a contained space that prevents code from affecting other systems). This enables teams to perform actions like synchronizing their data with Notion, creating tailored tools, and activating workflows through webhooks — automated prompts that initiate actions whenever a specific event occurs in another application — without having to depend on external systems.

You might not even need to write the code yourself. The company highlights that your chosen AI coding assistant can handle that for you.

The Workers will operate on the same credit system as Custom Agents, but Notion is offering this at no cost until August, encouraging developers to explore.

Synchronizing external data sources is also included in the Notion Developer Platform. Driven by Workers, the database synchronization feature can retrieve data from any API-enabled database. This means you could integrate information from resources like Salesforce, Zendesk, Postgres, and others directly into your Notion databases — ensuring the data remains up-to-date.

Zhao emphasized that this allows Notion users to “utilize your Notion database as a versatile canvas to drive both your workflows and your agents.”

Workers are also capable of creating agent tools with custom logic for instances when third-party integration via MCP — short for Model Context Protocol, an emerging standard for connecting AI tools to external data and services — is insufficient.

Additionally, a new feature permits Notion users to interact directly with external AI agents they utilize, assign tasks to them, and monitor their progress, akin to Notion’s own custom agents. At launch, Notion confirms compatibility with partner agents such as Claude Code, Cursor, Codex, and Decagon, with plans to incorporate more in the future.

An External Agent API is also available for teams wishing to link their internal agents with Notion, particularly those tailored for their specific organizational requirements.

Developers and agents engage with Notion’s innovative Developer Platform via the Notion CLI, a command-line interface designed for developers, accessible through the company’s Business and Enterprise Plans.

The Developer Platform signifies a strategic pivot for Notion, evolving it into a more programmable framework rather than just an application, positioning it to compete against other workflow automation services. As organizations increasingly seek to automate knowledge work and construct internal AI systems, a platform consolidating agents, custom code, and real-time data begins to resemble core infrastructure rather than just a productivity tool.

This direction aligns with the broader movement among AI firms, which have been advancing beyond chatbot functionalities to provide agentic tools capable of executing actions across varied software platforms.

“Any data, any tool, any agent — that’s the overarching vision for the Notion Developer Platform,” Zhao concluded.

As the tech world concentrates solely on AI models, Anthropic is experiencing an extraordinarily prosperous year.

The firm is on the verge of outpacing its primary rival, aiming to gather tens of billions in a funding cycle which could elevate its valuation to around $950 billion (OpenAI was valued at $854 billion during its March funding round), with business clients increasingly favoring Claude over ChatGPT. A recent study indicated that Anthropic has surpassed OpenAI with business clients, quadrupling its market share since May 2025.

Cat Wu, who leads the product team for Claude Code and Cowork at Anthropic, has played a pivotal role in that success. Since becoming part of the company in August 2024, Wu has guided Claude through a significant phase, enhancing it from a purely informational chatbot to a more advanced coding tool. Wu, who manages the creation of new functionalities, often collaborates with Boris Cherny, a vital member of Anthropic’s technical team and the developer of Claude Code, leading them to be dubbed Anthropic’s “Batman and Robin.”

Wu met with me during last week’s second annual Code with Claude conference in San Francisco, where she shared her insights on product strategy and her vision for how the use of Claude will evolve in the future. 

This interview has been condensed for brevity and clarity.

When considering product strategy, how much of it is influenced by your peers or rivals? Is that a concern for you?

Our primary focus is on remaining at the cutting edge, so I believe we instill in our team the understanding that AI will consistently improve. For us, the goal is to remain at this forefront. We do not concentrate on competitors. I think if you start considering competitors, you risk being perpetually a few weeks, or even a month, behind your capacity to execute. Therefore, this approach usually doesn’t keep you at the forefront.

Anthropic introduced at least six models last year and has nearly matched that number this year. Do you foresee this pace of development persisting?

We hope it persists (laughs). I believe the models continue to advance at a steady rate, allowing us to share them with our users consistently. The deployment may vary—like our approach with Glasswing—but as much as feasible, we want this intelligence to be advantageous to as many individuals as possible, and it must be managed very safely, which is why we approached Glasswing [the way we did].

[Glasswing is an initiative that Anthropic launched in April, inviting a select group of partner organizations—including firms like Amazon, Apple, CrowdStrike, and Microsoft—to access its new cybersecurity model, Mythos. Unlike many of Anthropic’s other AI models, Mythos is not set for a general public release. The company has expressed concerns that the model—designed to scan codebases for software vulnerabilities—might be misused by malicious actors.]

You previously mentioned that the future of work involves staff overseeing fleets of agents. This could eventually lead to situations where agents perform jobs better than humans, right?

Managing agents is incredibly challenging if you lack the skills for the job. Managers still need to be experts in their field. It’s a novel skill set that many will need to acquire, but managing agents is actually quite similar to managing people, in that you need to grasp why an agent made a mistake. Did it misunderstand my instructions? Was my request ambiguous? You need the ability to troubleshoot.

It appears that the long-term vision is to reduce team size. If agents can handle jobs, then does an intern become unnecessary?

Ideally, the notion is that everyone can achieve much more. For every role, there’s always a portion that is quite tedious. For me, that’s dealing with emails. Everyone has some aspect of their life similar to this… So, I hope that the AI agents can take on those tasks, freeing up time for everyone to pursue more exciting projects [in their free time].

What are you most enthusiastic about in the upcoming six months?

I believe the next significant development is proactivity. Last year was focused on synchronous development. Currently, people are transitioning to routines, like automating responses to customer support inquiries. I think the next phase is that Claude will understand your work patterns and autonomously set up some of these automations for you.

The malware research organization vx-underground, claiming to have the most extensive collection of malware source code, announced in a post on X that its data archive is approximately 30 terabytes.

In a response, Bernardo Quintero, the founder of VirusTotal, a web-based service that scans files for malware using various antivirus engines simultaneously, mentioned that his platform has around 31 petabytes of malware samples contributed by users so far. (One petabyte is roughly 1,000 times larger than one terabyte.)

In both instances, that’s a substantial quantity of data. To provide context, cybersecurity companies, artificial intelligence researchers, and threat intelligence firms regard collections like these as essential for training detection models and comprehending the evolution of attacks. This led us to question: How would these vast datasets actually appear when stacked as hard drives, both vertically and horizontally? And how would they stack up against something like the Eiffel Tower?

Someone in our news team posed this question to an AI chatbot, which provided a remarkably incorrect response.

Instead, we performed some basic calculations to estimate how tall these data reserves would be. Since both vx-underground and VirusTotal report “about” this much data each, “about” suffices for us in this scenario. 

Assuming we utilize 1 terabyte internal hard drives, as these are typically manufactured to have the same physical dimensions to fit in any computer. These standardized 3.5-inch internal hard drives measure 1 inch in height, which is precisely what we need to know for stacking purposes.

We are also assuming that the hard drives in this example are exactly 1 terabyte, as the actual usable capacity of a hard drive is usually slightly less. 

Using an online conversion tool, it appears that vx-underground’s 30 terabytes of malware data could fill 30 hard drives stacked together, reaching a height of 30 inches, or about 2.5 feet.

For perspective, this reporter stands at 6 feet tall. (Refer to the visual below, and yes, I acknowledge the poor operational security.)

By the same reasoning, VirusTotal’s 31 petabytes of collected data would require 31,744 hard drives, which when stacked, would reach approximately 2,645 feet high.

The Burj Khalifa in Dubai, the tallest building in the world, is a bit taller at 2,722 feet.

The Eiffel Tower stands at 1,083 feet tall. Therefore, VirusTotal possesses data equivalent to approximately two and a half Eiffel Towers.