PayPal reportedly rejected a takeover offer last month over concerns it was being undervalued.

Plans for a joint acquisition of PayPal have fallen through, according to Bloomberg News, weeks after the company’s board reportedly rejected a $53bn-plus offer from Stripe and US private equity firm Advent International.

Talks of a potential Stripe acquisition of PayPal have swirled for months. According to reports last month, Stripe and Advent offered to buy PayPal at $60.50 per share, which would have valued the fintech at more than $53bn and represented a 28pc premium to its share price in mid-July.

That offer was, however, reportedly rejected by PayPal’s board over concerns that it undervalued the company. Sources told Bloomberg that Stripe and Advent have now abandoned talks.

Founded in the late 1990s, PayPal has struggled in recent years to modernise against emerging rivals in the payments space such as Apple, Google and Stripe. PayPal’s shares have dropped dramatically over recent years as a result, but picked up somewhat last month after promising reports of a takeover by the $159bn Stripe.

PayPal’s shares have jumped nearly 10pc over the last month, now giving the company a market capitalisation of $52.5bn – up from roughly $43bn in February, when Bloomberg first reported on a potential acquisition. The fintech also recorded a 5pc growth in net revenue in the second quarter.

PayPal and Stripe would each benefit from the possible acquisition, according to Chris Jones, managing director at PSE Consulting.

“PayPal’s wallet could build on the early success of Link, Stripe’s consumer-facing accelerated checkout, which already counts more than 200m consumer accounts, and would create further opportunities to exploit Stripe’s $1.1bn investment in stablecoin infrastructure through its purchase of Bridge,” he told SiliconRepublic.com last month.

The US fintech brought in HP’s Enrique Lores as its new CEO earlier this year, replacing Alex Chriss, whose work was “not in line” with board expectations. By June, PayPal had reportedly planned to shut down its venture capital arm.

Stripe, meanwhile, is investing heavily in product development and making strategic acquisitions, including the programmable wallet company Privy, stablecoin orchestration platform Bridge and billing platform Metronome.

It recently announced that it is acquiring large language model marketplace OpenRouter at a reported cost of between $7bn and $10bn as part of efforts to help businesses optimise token costs.

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.

In a nutshell: Cloudflare has cut the memory used by the DNS cache behind its 1.1.1.1 resolver by more than half, freeing about 100TB of RAM across its network. The company did not replace server hardware or remove memory modules. Instead, engineers changed how cached DNS data is stored in software.

The optimization work was centered on Big Pineapple, Cloudflare’s DNS caching platform behind the popular 1.1.1.1 resolver, DNS Firewall, and several other DNS services they offer. In a technical blog post, systems engineer Sebastiaan Neuteboom detailed five changes to the Rust codebase that shrank a typical cache entry from 953 bytes to 420 bytes.

The changes also made the cache faster. Cloudflare reported that insert throughput rose from 625,000 to 893,000 entries per second, while lookup latency fell from 828 nanoseconds to 670 nanoseconds. During the rollout, from mid-May to early July, p99 memory use per instance dropped from 9.3GB to 5.3GB.

Cloudflare says Big Pineapple holds more than 250 billion DNS entries at any given time, and at that scale even small amounts of wasted memory add up fast. The company estimates that a single unnecessary byte in each cache entry costs more than 250GB across its fleet.

Big Pineapple, Cloudflare’s DNS caching platform, was redesigned in Rust to free about 100TB of RAM.

One of the biggest changes involved Rust’s Vec and String types. Those containers are designed to grow, so they store capacity data in addition to the data itself. But DNS records do not need to expand after they enter the cache. Cloudflare replaced those structures with fixed-size boxed slices, eliminating metadata that was no longer useful. The company said that change alone saved more than 15TB of RAM.

Engineers also rethought how DNS responses are laid out in memory. Big Pineapple used to store three separate record lists per response. The new design puts the records in one buffer and identifies them with two-byte offsets.

The cache no longer retains owner names when they simply repeat the queried domain. Cloudflare rebuilds those names when it serves the response. It also now stores record data as length-prefixed raw wire-format bytes. That change addresses a mismatch in the earlier design. A four-byte IPv4 A record occupied the same 144 bytes as the largest DNS record type Cloudflare caches, the relatively uncommon NAPTR record. Storing raw data in a variable-length format reduces that waste.

The 100TB of RAM Cloudflare recovered across its fleet is roughly equal to the memory in 130 of its Gen 13 servers, each of which carries 768GB of DDR5-6400. The company had considered a 1,152GB configuration for that server generation but passed on it after weighing the cost of higher-capacity RAM.

With the RAM savings, Cloudflare plans to allocate the memory to larger DNS caches. Keeping more records locally should improve cache hit rates and reduce the traffic it sends to authoritative DNS servers.

This project follows another major memory-efficiency effort completed last year, when Cloudflare rewrote its FL2 request-handling layer in Rust. That’s probably a good way to spend resources and squeeze more out of its existing infrastructure as server memory gets pricier.

Apart from the new chipset, the camera upgrades, and Gemini Intelligence, the Pixel 11 Pro shipped with something that turned a lot of heads, in enthusiasm no less: the RGB light module on the back called HiLight. 

Before Google announced the feature, I was actually excited about it, thinking about all the potential use cases; the light could have been a subtle notification queue for users, one that doesn’t compel them to unlock the phone right away. However, Google tied the hardware to two use cases: Gemini interactions and calls from favorite contacts.

That is exactly why tech content creator @Dhananjay_Tech (DhananjayBhosale) created an app called HiLight Studio, with a simple pitch: unlock HiLight’s true potential by opening it up to third-party apps and letting you have some fun with it. After spending a few days with the app on a Pixel 11 Pro XL, I have some thoughts, both about what the HiLight is actually capable of, and how much Google actually left on the table with it.

That’s where HiLight Studio comes in

On the home page, you can try out plenty of HiLight effects. The Rainbow effect comes closest to what we’ve seen in the promotional materials so far, with the modules lighting up in a rotating gradient. You also get five other effects to choose from: Random, Comet, Pulse, Breathe, and Wave. While this page is called Live (designated at the bottom of the screen), the Style page to the right lets you enable HiLight as an always-on light. 

You can also choose different ways you want the LED modules to light up, including a gradient finish (with the option to select the start/end color), a breathing look (with the option to set the timing per cycle), and more. 

There’s an Apps section as well, wherein you can customize how the HiLight module glows up upon receiving a notification from an app. In my case, I’ve set a Chasing green light for WhatsApp, set as “on notification” rather than “while open.” Similarly, you can also create per-app rules for Instagram, Messenger, Teams, Gmail, or Slack.

This is where the hardware finally makes sense

This is one of the most intuitive and natural use cases of HiLight in my opinion, something that I’d like to remain this way. For what it’s worth, the app genuinely makes Google’s HiLight module interesting. It turns the boring system implementation into something that I always imagined it would be: a glanceable notification system.

Here’s the catch: enabling wireless debugging locks me out of the banking and payment apps I use daily, for security reasons that are frankly fair, not paranoid. So even though I like what HiLight Studio does, I can’t run it full-time. Second, and more importantly, actually getting the app working is a genuinely arduous process for a regular user.

Getting to what Google left behind is more tedious than it should be

You need two apps installed: HiLight Studio and Shizuku. First, enable Developer Options by tapping the build number seven times under “About phone,” because apparently that’s still how we’re gatekeeping settings in 2026. Then enable wireless debugging, install Shizuku, pair it using the six-digit code from the notification, and tap “Start.” 

Shizuku, for the uninitiated, lets third-party apps borrow first-party system permissions, which in this case means the HiLight module. It’s only on the Play Store for Android 16 devices, so it wasn’t even available for my Pixel 11 Pro XL running Android 17, forcing a sideload instead. However, some of you might already know that the app was recently exploited by RedHook malware, which would also keep most users from installing it.

From there, open HiLight Studio, head to Setup, and under “Privileged access,” select Shizuku. Grant notification access too. Get everything right, and you’ll see a small “8 LEDs · Shizuku” confirmation appear, followed by one final toggle next to “System has HiLight” that unlocks the effects at the bottom. 

Getting here is easier said than done

You see? It took me around 150 words to explain how the process works, and it would probably take you around 10 to 15 minutes to set everything up, provided you’re familiar with Developer Options and apps like Shizuku.

The app works about 70 to 80% of the time, but sometimes a WhatsApp notification doesn’t trigger the effect. Further, rebooting the phone disrupts the connection between HiLight Studio and Shizuku, and you have to re-establish it to use the lights.

I’d have to give up access to the payment apps I use every day, keep wireless debugging enabled, and rely on a third-party bridge that might or might not be a security issue, all to unlock functionality that Google could have offered natively. These aren’t things that a regular user would want to do on their smartphone.

For me, HiLight Studio proves what Google is missing

I really appreciate the creator of the app, as it lets users explore HiLight’s true potential. However, a third-party app can’t beat a native, system-level implementation that would make the feature much more useful and worthwhile for everyone, not just enthusiasts.

The app simply proves that the hardware can do much more than Google currently allows, and that’s the key takeaway. If someone at Google is reading this, I genuinely hope you’re planning to add some fun HiLight functionality in one of the upcoming Pixel Drops, because otherwise, this feels like a missed opportunity.

It used to be that Vijay Pande was better known in academic circles than investor circles. That changed pretty abruptly a dozen years ago, when Marc Andreessen and Ben Horowitz — who’d spent their firm’s first five years explicitly avoiding healthcare and life sciences — decided the category was worth betting on after all and handed the keys to Pande. At the time, he was a Stanford chemistry professor who was best known for building Folding@home, the distributed-computing project that turned millions of home PCs into a supercomputer for disease research. Over the next decade-plus, he grew a16z’s bet into a practice managing close to $4 billion.

So it was somewhat unexpected when in June of last year, Pande walked away from it all to start something much smaller. In fact, his new firm, VZVC, co-founded with longtime investor Zach Werner, is built around a handful of concentrated bets a year rather than dozens, it has no associates, and it relies heavily on AI for its day-to-day operations.

To learn more about Pande’s hard pivot, we talked with him this week about why he’s making just a handful of concentrated bets rather than spreading himself thin in the current market — and about one of the more interesting conundrums in AI-driven biotech: unlike text, biological data can’t be scraped off the internet, so nearly every company ends up building its own walled-off dataset. What does that mean for all the advances AI in medicine has promised, and who actually gets access to them?

This conversation has been edited for length and clarity. You can also listen to the fuller conversation (below).

You’ve said biology is moving from a “science of discovery” to something you can engineer. What does that mean?

For a lot of the way drugs have been developed, there was very much a fortuitous aspect to it. I think what’s shifted is that AI and machine learning allow computers to wrap their type of understanding around something very, very complicated… to try to figure out what targets you want your drugs to hit, for specific diseases, to be able to make those drugs, and now even to help in the clinical trials — which are the most expensive part of the process.

I thought clinical trials were getting cheaper because drug developers are using more synthetic data, so not as many people are needed for these trials.

That’s, I think, very much an aspiration.

The cost and time to get to clinical trials has been shrinking, especially with AI, but it could still cost hundreds of millions of dollars to run a trial, which is why drugs are very expensive. The probability of a drug going successfully from the first trial to the end of the third trial is just 20%. If 8 out of 10 fail, and these things cost hundreds of millions of dollars, the amortized cost gets really high. The reason they fail typically is not that the biologist did something wrong; it’s that all the experiments these drugs were designed on were on animal models like mice, and in the end, animal models are just not very predictive of humans. The AI model is not going to be perfect, but it’s going to be way better than any animal model would be, and once it crosses that bar, that’s where it gets really exciting.

[The phase after that is]: Is the drug the right drug for me?

You mean personalized medicine. . .

The jargon here is so-called precision medicine. If you go to a doctor with something not trivial, they have to guess what’s going on, because there’s only so much they can tell. Then they give you a drug — and if that doesn’t work, they give you another drug, then another drug. This happens in cancer, it happens in lots of different areas. We would all be much better off if the first drug was the right one. Typically, your blood test values are compared to population averages. But really, they should be compared to: is this [result] weird for you? What we’re starting to do also on the medicine side is [the ability] to just understand what would be right for the individual.

Would you say the path to this moment has been slow and steady, or did it spike more recently?

I think it’s lots of different things [coming together]. So for instance, precision medicine for the longest time was based on genomics. But the reality is your genome is kind of like the blueprint for your house on day one, but your house is fairly different now compared with the moment it was built. So there are many other things that people can now measure in proteomics and so on that are much more relevant for understanding disease and where your body is now. There has also been [a lot of] automation in robotic measurements that is naturally tied into AI, and those two go hand in hand really well. 

Over the last decade, there’s been this steady clip fforin both AI for biology and AI for chemistry. The biology part is like, how can we treat this disease? And then the chemistry part is, how can we come up with a drug to go after that specific protein? There have actually been very significant advances over those 10 years. 

You mentioned that biology is one of the few places AI can’t just scrape data off the internet. What does that mean for how the field develops?

It’s a place where you don’t have any of this data that people can just all train the same thing, and your data can’t be distilled from one model to another. It’s a really interesting play from just the pure AI sense.

Doesn’t that echo a familiar problem in medicine, though — doctors operating in [territorial, often competitive] silos?

You’re onto something really big here. Let’s say [someone] has some type of cancer, and it’s both an issue in oncology and endocrinology — those two doctors really don’t sync together very well. What is really intriguing about AI is that it can, in principle, be a specialist in everything, and it can start to see things that really any single human being couldn’t. It would be equivalent to having a team of the very best doctors all clamoring together in that moment.

But is there enough data sharing for that vision to actually be realized? I understand why founders and investors want to protect their [respective findings], but . . .

I think one of the bigger trends is that we’re starting to see a shift toward building these atlases of biological information — which, from a technology standpoint, are typically foundation models. And as they become more common, I think we’ll see the same thing that’s happened with open-source LLMs, which do very well against the corporate ones: open-source foundation models in biology having a very broad impact.

You’re involved with Genesis Therapeutics, which came out of your lab at Stanford, and Insitro, the drug-discovery company launched by Daphne Koller, a former colleague at Stanford. You say you’re also incubating a company with a founder you’ve known for 20 years. What are you looking for in founders, and in what areas?

There are two areas that I’ve been spending most of my time on. One is AI for healthcare delivery, which I did a ton at a16z as well, and then AI for clinical trials.

One of the things that’s most important to me [about founders] is that we can really trust each other — founders that have high integrity, that do what they say they’re gonna do… I’m expecting this relationship to be 5, 10 years plus into, ideally, their next company. I want to work with people who are thinking long term like that. Ideally, these are people who are not just trying to win and beat other people, but really thinking about the question: how do we win together?

What have you gotten right and wrong in your investing career so far?

When I started talking about AI and machine learning and technology and medicine and bio 10 plus years ago, there was a lot of resistance and a lot of people saying, ‘Oh, that’s never going to happen. That’s never going to be useful,’ and so on. That resistance is largely gone and seeing this arc is very fulfilling.

I think it took me some time to really appreciate that as seductive as the coolest technologies are, it really always comes back to go-to-market. I tell my founders, especially the ones who are coming from the science or the product side, for them to take all their brilliance and creativity and really apply it to the go-to-market side, that the go-to-market part is at least as hard or harder than the technology side.

Help us understand how you’re designing this new firm differently, compared with what you were running at a16z.

Right now, we’re doing something really quite different… VZ is named after me, Vijay, and my co-founder, Zach Werner — he’s the Z. We’re intentionally really quite small… on the investment side, it’s really just the two of us. We were actually intending on hiring associates, but it turned out, with the agents that we’ve built up, not to be something that we need to do.

How concentrated is “concentrated”?

We’re [not] driving 30 bets per year… we’re talking about probably five, not a lot of investments — very concentrated. Adding a company at a typical fund is like adding a Facebook friend — that’s something you do pretty quickly. For Zach and I, it’s more like . . . wanting to have another child. This is a big deal for us.

With that structure, who are you competing against for deals?

The funny thing about this model is that typically we’re not trying to compete for a hot round — people make room for us. It’s a very different thing than trying to get the hot Series A or Series B. Largely, people want us as investors because of what Zach and I can do, and how hands-on we can be. When I look at people who are inspirations, I look at someone like Antonio Gracias at Valor — he’s well-known now because of the SpaceX deal, but he’s been doing what he’s been doing for 20 years. What Thrive has done, with a more concentrated portfolio, is also a real inspiration. Obviously, a16z is sort of in my DNA as well, but I think those other ones are new additions to how we think about things.

What’s overhyped right now in AI and biotech?

The reality is that AI can find insights that we can’t get from just humans alone. The thing that always gets tricky is when there’s this call that AI is going to cure all everything. The reason for hesitance there is not because of any doubt about AI — it’s about doubt of the data. LLMs work because there’s so much data to learn from. When the data is just simply not there, then AI can’t magically solve that problem.

Anthropic is permanently increasing Claude Code’s standard weekly usage limits by 25% for Pro, Max, Team, and seat-based Enterprise plans, but it’s not as good as it sounds.

If you use Claude, you’re actually getting a 17% reduction compared to what you have today.

Claude Code currently has a temporary 50% increase in weekly limits, which Anthropic says will remain in place until September 14.

“Starting September 14, we’re permanently raising standard weekly limits in Claude Code by 25% for Pro, Max, Team, and seat-based Enterprise plans,” Anthropic wrote on X. “Until then, the current 50% increase will be in place.”

It’s a clever way to frame words, as it almost sounds like you’re winning as a customer, but you’re not, and it’s a downgrade.

If Claude Code’s original weekly allowance was 100, the temporary boost gives you 150 today. On September 14, that drops to 125.

In other words, you will have 25% more Claude Code usage than you did before the temporary promotion, but about 17% less than you have right now.

Anthropic says more Claude Code usage changes are coming

Anthropic later deleted the original thread and posted a clarification, where it clearly admits the reduction.

“Compared to today, this works out to a 17% reduction in weekly limits on Claude Code,” the company said.

“We’re working on exciting changes that will make it feel like you’re getting more from Claude, while having more visibility and control of your usage. Can’t wait to share them.”

Anthropic says Claude usage can vary based on factors including conversation length, model choice, tool usage, and effort level, so the weekly allowance does not translate into a fixed number of Claude Code prompts.

The current 50% temporary increase remains available through September 13, with the new permanent limits taking effect on September 14.

Overall prevention scores can hide what happens after initial access. Once attackers are using valid credentials, prevention drops sharply.

The Blue Report 2026 measures defenses technique by technique across 338 million simulations run in customer production environments.

Advertisement

Get the report