Firebird, a US-based AI cloud company, opened what NVIDIA calls the CIS region’s largest AI factory in Hrazdan, Armenia on Saturday. Armenian Prime Minister Nikol Pashinyan attended, alongside Kazakhstan’s deputy prime minister Zhaslan Madiyev and David Allen, the US chargé d’affaires in Armenia.

That guest list is the story. A US company opened a facility in a country inside Russia’s traditional sphere of influence, with an American diplomat and a Kazakh minister watching.

The licence is the load-bearing fact

None of this happens without export approval. Firebird secured a US export licence in November 2025, the regulatory step that allowed advanced NVIDIA hardware into Armenia at all.

The politics were made explicit in February. Firebird and the US government jointly announced Phase 2 at a Yerevan press conference during Vice President JD Vance’s visit, scaling the project to $4 billion and roughly 50,000 GPUs with licensing for a further 41,000 NVIDIA GB300 units.

Phase One was $500 million. The announcement claimed Armenia would host one of the world’s five largest AI GPU clusters.

What is actually running

The opened flagship, DC-1, is specified at 6,144 NVIDIA B200 GPUs across 15 megawatts, liquid-cooled and built on NVIDIA’s reference architecture. The target is more than 70,000 Rubin and Blackwell GPUs and 300 megawatts by the end of 2027.

That is a twentyfold increase in power capacity in roughly 17 months. A planned DC-2 is listed at 75,000 VR200 GPUs across 125 megawatts, with a third site at similar scale.

The delivery speed is genuinely notable. NVIDIA says the site went from plan to operating capacity in just over six months, with Schneider Electric supplying switchgear, three-phase UPS systems and rack enclosures, and Vertiv providing chilled-water cooling.

The power trick

The facility runs on NVIDIA’s DSX platform, which codesigns compute, networking, power and cooling as a single system rather than separate procurements. NVIDIA claims it allows up to 40% more GPUs on the same footprint by recovering stranded power.

For a site scaling from 15 to 300 megawatts, that coordination is where the capacity maths gets made. It is also, in NVIDIA’s own framing, a tokens-per-dollar argument rather than a performance one.

Who is paying for it

Firebird said NVIDIA intends to invest in the company, following an earlier CoreWeave investment this year. The pattern is familiar: NVIDIA takes a position in a company that then buys NVIDIA hardware at scale.

NVIDIA has now committed more than $40 billion to AI equity positions in 2026, and the structures have grown more creative. It took a $2.1 billion warrant in IREN as part of a five-gigawatt data centre deal.

The same logic runs through its sovereign deals. NVIDIA invested $1 billion in Naver while striking a $500 billion arrangement with SK Group, an approach critics have labelled circular financing.

The neocloud playbook

Firebird fits a category that barely existed three years ago: companies that buy GPUs at scale and rent them to model builders. CoreWeave, an early Firebird backer, has signed a multi-year deal with Anthropic to run Claude at production scale.

Europe has its own version. Nscale committed €695 million to Portugal with Microsoft, reaching a $14.6 billion valuation in two years, on much the same thesis about sovereign capacity.

Firebird’s first named customer is Perplexity, which is using the site for its answer engine and AI agent platform.

Where this goes next

Firebird is pursuing an approximately two-gigawatt roadmap spanning Armenia, Kazakhstan and further markets, which explains why a Kazakh deputy prime minister was at an Armenian ribbon-cutting.

The strategic logic is straightforward. Compute is becoming the thing countries align around, and the United States has found a way to place it in a region where its influence has historically been thin.

The open question is demand. Building 300 megawatts in the Caucasus assumes customers will route workloads there rather than to cheaper or closer capacity, and one answer engine does not settle that.

Lofree Flow 2

MSRP $159.00

Advertisement

“The Lofree Flow 2 types beautifully, but its awkward extras lack the same polish.”

Pros

• Excellent typing experience
• Premium aluminum construction
• Smooth, refined Surfer switches
• Hot-swappable design
• Wired and wireless connectivity

Cons

• Touch strip activates accidentally
• Awkward side-mounted USB-C port
• No storage for the wireless receiver
• VIA setup is needlessly difficult

“Why you can trust Digital Trends – We have a 20-year history of testing, reviewing, and rating products, services and apps to help you make a sound buying decision. Find out more about how we test and score products.“

The Lofree Flow 2 made an excellent first impression. Its aluminum body feels reassuringly solid, its minimalist design looks expensive without screaming for attention, and once I started typing, my fingers felt strangely lithe.

That isn’t a word I normally associate with fingers, but it fits. Moving between keys required less effort without giving me the shallow, slightly dead sensation I usually get from laptop keyboards. I even felt like I was typing faster.

For once, the numbers supported the new-product placebo. I normally average between 90 and 95 words per minute on Monkeytype, but I regularly crossed 100 words per minute with the Flow 2. That hardly qualifies me for competitive typing, assuming such a terrifying sport exists, but the difference was consistent enough to notice.

Living with the Flow 2 also revealed several decisions that aren’t nearly as polished as the keyboard itself. None made me want to stop using it, but they suggest Lofree spent more time perfecting what happens under your fingers than considering everything your hands might do around it.

Price and configuration

I tested the latest 84-key Mac version, which costs $159 and comes with Lofree’s Surfer linear switches. These belong to the newer Cloud Series developed for the Flow 2, replacing the Ghost and Phantom switches offered with the original Flow.

Lofree also sells the keyboard with quieter Void linear switches and tactile Pulse switches, so the exact typing experience will depend on the configuration.

At this price, the Flow 2 competes with premium low-profile keyboards from Keychron, NuPhy and Logitech. That makes its smaller frustrations harder to dismiss. A less expensive keyboard can get away with asking its owner to improvise. A premium one should have already thought through where the dongle goes.

The Flow 2 offers an excellent typing experience

The Flow 2 sits somewhere between a traditional mechanical keyboard and a laptop keyboard without feeling like a compromised version of either.

Its keys have enough travel to make each press feel deliberate, while the low-profile design keeps movements short. Each press lands softly without becoming mushy, the Surfer switches are smooth and consistent, and the wider keycaps give my fingers a comfortable target.

My regular mechanical keyboard feels heavier by comparison. I encounter less resistance while moving through a sentence, which probably explains those higher Monkeytype results. That’s especially pleasant during long stretches of writing, when a heavier keyboard can gradually make every key feel like a tiny administrative task.

The aluminum construction completes the effect. The board doesn’t noticeably flex, and it stays planted on my desk. Its restrained design also avoids the usual gamer aesthetic. You get white backlighting, but no rainbow running beneath the keys like a tiny nightclub.

My space bar has a faint springiness if I listen for it, although it’s negligible during normal typing. I didn’t encounter the prominent metallic resonance reported in some other reviews. Whether that comes down to the Surfer switches, production differences or plain unit-to-unit variation is difficult to say.

The Flow 2 is hot-swappable, so you can replace its switches without soldering. That gives enthusiasts some room to alter the typing experience and means one failed switch doesn’t condemn the entire keyboard.

I’d still approach that feature cautiously. The plastic clips on low-profile switches seem more delicate than their chunky full-height equivalents, and compatibility is narrower than it is in the standard MX-style market. Hot-swappable doesn’t mean every mechanical switch you find online will fit.

Fortunately, I haven’t wanted to replace the stock switches. People who normally open a new keyboard and immediately start adding foam and tape in pursuit of the perfect “thock” may be surprised by how refined this one already sounds and feels.

The touch strip creates some daily frustration

The touch-sensitive strip on the right side is the Flow 2’s most conspicuous experiment. It controls volume and can adjust screen brightness. When I remember it exists, sliding a finger across it feels more elegant than repeatedly tapping a function key.

The trouble is remembering it at the right time while somehow forgetting it whenever I touch the keyboard’s edge.

Even after getting used to the Flow 2, I still trigger the strip by accident. My hand occasionally brushes against it, particularly when I’m repositioning the keyboard. Physical buttons or a dial would have been less novel, but they would also have been easier to use deliberately.

The USB-C port occupies that same side. Depending on how your desk is arranged, the cable can stretch toward the mouse area instead of disappearing neatly behind the keyboard. It works perfectly well, but its position seems better suited to the keyboard’s internal layout than anyone’s actual desk.

The 2.4GHz receiver, meanwhile, has nowhere to live. I keep mine in the box because I know exactly how this story ends otherwise. It’ll sit loose on my desk, migrate into a drawer and eventually disappear into the same dimension that collects missing socks and single wireless earbuds.

That omission feels especially strange on a compact wireless keyboard. The Flow 2 is small enough to carry between desks, yet its smallest and easiest component to lose has no home inside it.

My most personal complaint involves the mode selector underneath the keyboard, which switches between wired, wireless and off. Hiding it there prevents accidental changes and would normally make sense, except I also have cats.

When one jumps onto my desk and decides the Flow 2 is a perfectly acceptable walking surface, I can’t quickly disable it without lifting the entire keyboard. A side-mounted control would make those little hostile takeovers easier to manage. This may be a specific complaint, but pet owners understand that every exposed surface eventually becomes a collaborative feature.

The 84-key model has a 3,000mAh battery rated for up to 120 hours with the backlight switched off. That’s competitive for this category, although maximum-brightness backlighting cuts the advertised figure substantially. Since the Flow 2 spent most of its time on my desk, battery life never became the central issue that its other design choices did.

VIA customization is harder than it should be

The Flow 2 supports VIA, which lets you remap keys and configure different layers through a browser interface. It should be one of the keyboard’s most useful features, especially for anyone who regularly moves between operating systems.

Actually getting it working was far harder than necessary.

During my testing, I couldn’t find the correct JSON configuration file on Lofree’s main website, download center or support pages. VIA wouldn’t properly recognize the keyboard without it, so I spent the better part of a day searching before contacting Lofree. The company eventually sent me the file directly.

Lofree now provides some Flow 2 configurator resources through individual regional and product pages, but they remain surprisingly difficult to locate through its central download page. A keyboard advertised as supporting VIA shouldn’t require buyers to know which obscure corner of the manufacturer’s website contains the missing ingredient.

VIA works well enough once configured, although the Fn key still can’t be customized. I wanted to reconfigure it while using the Flow 2 with Windows, but that limitation made switching layouts less flexible than expected.

I eventually kept the keyboard connected mostly to my Mac, where the native layout made sense and I could concentrate on writing instead of managing the thing I was writing with.

Should you buy the Lofree Flow 2?

The Flow 2 is worth buying if typing quality, construction and appearance matter more to you than seamless customization. It feels refined immediately, so enthusiasts may be surprised by how little they want to modify. More casual buyers can get an excellent mechanical typing experience without surrendering a weekend to foam, tape and switch research.

It’s less convincing if you frequently switch operating systems, expect VIA to work without a scavenger hunt or regularly carry the receiver between desks. Mac users who intend to leave it in one place will encounter fewer of its rough edges.

Anyone indifferent to the precise woodland noise made by a space bar may simply see an attractive aluminum keyboard that pairs nicely with the MacBook Neo colors.

I still occasionally trigger the touch strip, and the receiver remains in protective custody inside the box. Yet I keep returning to the Flow 2. I thought it merely made me feel faster, but Monkeytype says otherwise. Apparently, the keyboard tested my patience and improved my typing speed at the same time.

Framework, a company best known for producing modular, highly repairable and customizable PCs and laptops, has suffered a large data breach, which has compromised customer information, according to multiple media outlets and an email from the company that has been shared online. The stolen information includes customer names, email addresses, phone numbers, login IPs and physical addresses. In the emails some customers have received, the company says that order and payment information wasn’t accessed in the breach.

A Framework representative didn’t immediately respond to a request for comment.

According to TechCrunch, Framework spokesperson Eric Schumacher said that all customers were affected, without specifying an exact number.

The breach appears to have been caused by a Metabase Cloud vulnerability using a zero-day exploit, a previously unknown security flaw in the software. In a blog post, Metabase says the vulnerability has already been patched but warns that those self-hosting may still be vulnerable and should upgrade to the latest point release.

These types of breaches aren’t uncommon, even for large companies, and they hit across the industry. CNET reporting has shown that data breaches are becoming increasingly widespread and severe, while companies are often providing less transparency about the scope of the compromised information and the methods attackers use to gain access.

If you’ve been affected by the Framework data breach, there are some measures you can take to minimize the damage. Change any exposed and reused passwords, enable two-factor authentication if you haven’t already, keep a close eye on any payment cards linked to that account, and revoke access to services you no longer need.

Framework has been hit hard by the ongoing memory shortage and has raised prices twice this year. The cost increase for the memory modules was so significant that it actually pared down the RAM loadout on the Framework Laptop 13 Pro, despite the product already being offered for preorder. The latest breach may further challenge consumer confidence at a time when the company is already navigating pricing pressures and supply chain challenges.

Ajay has worked in tech journalism for over a decade as a reporter, analyst, product reviewer, and editor. He got his start in consumer tech, breaking Android news at Newsweek before going to PCMag, where he reviewed hundreds of smartphones, battery packs, and chargers as a Mobile Analyst. He also worked at Lifewire, a Dotdash Meredith brand, as a Tech Commerce Editor, putting together tested best-of lists and assigning product reviews across categories including smart home, uninterruptible power supplies, generators, and automotive tech.

Most recently, he was Section Editor, Mobile at Digital Trends, spearheading his team’s coverage of breaking news, features, reviews, roundups, deals, and more across a variety of mobile products, including phones, wearables, VR headsets, batteries, and chargers. If you want Ajay’s advice about anything tech, especially solar panels, UPS, batteries, EVs, and charging technology, you can reach him at ajkumar@cnet.com.

See full bio

Advertisement

I use ChatGPT pretty much every day, but somewhere along the way, I’ve also developed a habit of poking around and looking for features I haven’t tried yet. Some discoveries are genuinely useful, while others are the kind I’ll probably remember once every six months.

Every once in a while, though, I stumble across something that actually changes how I use ChatGPT. That happened recently with three features I’d somehow been overlooking. I’ve started using all three regularly now, and if you spend a lot of time in ChatGPT too, I think you’re missing out by not giving them a shot.

I let ChatGPT figure out my mornings

One feature I’ve started relying on a lot is scheduled tasks. ChatGPT can now run these for me automatically every day in my local time zone, which means I don’t have to remember to come back and ask it the same thing every morning. I’ve set mine up as a daily work brief. Since a lot of my work happens across Slack and Gmail, I connected both and asked ChatGPT to pull together the things that actually deserve my attention. Once everything is connected, it can read across those sources and immediately build a brief using what’s happening there right now. And this is where it becomes genuinely useful for me. Every morning, I get a short list of around three to five things that are most likely to need my attention that day. ChatGPT tells me what each one is about, why it matters, how urgent it seems, and what I should probably do next, along with a link or citation so I can jump straight to the original message.

Instead of opening Slack, then Gmail, and slowly figuring out what needs to happen first, I already have a pretty good idea of how my day looks. I can tackle anything urgent first and plan everything else around it. I’ve even asked ChatGPT to send the brief when absolutely nothing important is waiting for me. On those glorious mornings, all I get is a simple, “No urgent items found.” Quite frankly, that might be my favorite version of the brief. ChatGPT uses Codex to help set this workflow up, so you’ll need a Pro subscription to use it this way. Pro comes with plenty of other perks as well, and considering how much I use ChatGPT throughout my workday, I’ve found the subscription easy to justify. The whole idea reminds me quite a bit of the Daily Brief experience you can get with Gemini. The result isn’t dramatically different: both are trying to save you from manually digging through everything before your day has even properly started. But ChatGPT has become the AI tool I naturally reach for first, so having my morning brief live here fits my workflow better.

I moved in without unpacking everything again

I’d be lying if I said I’ve always been a ChatGPT person. Before making the switch, I relied pretty heavily on Claude, and for a long time, it simply fit the way I worked better. Naturally, that also meant Claude knew quite a bit about me and how I liked to get things done. That was actually my biggest hesitation when I decided to move over to ChatGPT. I wasn’t worried about learning a new interface or figuring out where everything lived. I was worried about starting from scratch. Thankfully, ChatGPT now has an Import from Claude feature, and it made that switch a lot less painful for me. Instead of manually rebuilding all that context conversation by conversation, I could bring my existing Claude data over and give ChatGPT a much better starting point.

For example, if Claude already knew how I preferred my research to be like, the kind of work I regularly do, or other details I’d previously asked it to remember, I didn’t have to sit down and feed all of that information to ChatGPT again. It might sound like a small convenience, but if you’ve spent months building up context with another AI assistant, you’ll know just how annoying starting with a completely blank slate can be. ChatGPT can also keep connected sources up to date by syncing new or changed content, so the information it can work with doesn’t necessarily stay frozen at the moment you connect it. For me, it meant switching AI tools without having to introduce myself all over again.

I can finally take my work with me

It gets even better from here because ChatGPT also lets me download the work I create inside it. This sounds like such a basic feature until you actually need to move something out of ChatGPT, and then you realize how useful it is. I run into this quite often with my study projects. If I’m working on something for my psychology thesis, for example, I might spend a long time inside ChatGPT brainstorming ideas, organizing my research, cleaning up notes, or turning a messy collection of thoughts into something I can actually work with. Once I’m happy with it, I can download the finished work as a Markdown file, PDF, or Microsoft Word document and continue working on it wherever I want.

That’s a much nicer experience than what I was doing before. I’d finish something in ChatGPT, open Word or another app, copy everything over, fix the formatting that inevitably went wonky somewhere along the way, and then continue working. It wasn’t difficult, but doing it repeatedly got very boring and time-consuming. Now, I can keep ChatGPT as the place where I brainstorm and build something, then simply download the result when I’m ready to take it elsewhere.

Your turn to give them a try

If you use ChatGPT regularly, I’d definitely recommend giving these features a try. I’ve been having a genuinely good time discovering new ways to use ChatGPT, and clearly, I have no intention of stopping anytime soon. So, if I stumble across another feature that makes my day a little easier, you can bet I’ll be back to tell you about it. Until then, give these three a shot. Your ChatGPT routine might just get a little better.

Apple quietly killed the ceramic Apple Watch in 2020, and for the fans of the fancy material, it has been an inexplicable absence ever since. For those catching up, ceramic was the lineup’s most premium material option. 

It was smooth to touch, scratch-resistant, and had that distinct non-metallic appearance, a characteristic of all titanium and aluminum frames. Now, according to Bloomberg’s Mark Gurman, Apple is actively planning its return.

Why does the ceramic case’s return actually matter?

While it isn’t mentioned clearly in the report, Apple could re-introduce a ceramic version of the Apple Watch with either the Series 12 in September or in 2027. “It will also offer new colors and bands, and a series of new health and fitness-related upgrades,” the report mentions.

For those catching up, ceramic is harder than stainless steel and retains its polish far longer. Basically, it’s the perfect blend of a luxury and a sporty finish. It’s the kind of materials upgrade that could serve as a major marketing campaign for the lineup without changing a single feature. 

For anyone who missed it the first time, or who has been wearing a titanium Ultra and doesn’t like the finish, the ceramic version could be the way to go. Fair warning: the Apple Watch could get fairly expensive with the ceramic case.

What else is Apple changing with the Series 12 and Ultra 4?

Bringing back ceramic is a smart way to restock the premium end of the Apple Watch lineup, but in my opinion, it will buy the company some time to figure out what the smartwatch will look like over the next decade: Apple’s yearning for a major Watch overhaul. 

Quite a bit, actually. Beyond the ceramic case option, both the Series 12 and the Ultra 4 are getting their first meaningful performance bump in years. The performance upgrade could be accompanied by new colors, updated band options, and a bunch of new health and fitness-related features.

Together, the performance upgrade on the Series 12 and Ultra 4 and the potential arrival of a ceramic case version suggest Apple wants to ship a comprehensive Watch lineup this year.

The takeaway: Memory prices have been trending upward for more than a year, with many analysts predicting further increases in the near future. A prominent GitHub developer has now published a series of graphs and charts showing just how quickly prices have surged to 2007 levels on a per-unit basis, effectively undoing two decades of progress in making PC hardware more affordable for consumers.

According to computer science researcher and software performance expert Daniel Lemire, skyrocketing memory prices are a “historical anomaly,” with the unprecedented increases effectively erasing decades of steady, exponential cost reductions.

He believes that for prices to normalize, either software developers will have to find ways to build AI systems that require less memory, or the hardware industry will need to increase production capacity.

Driven largely by rising demand from the AI sector, global memory prices have surged dramatically over the last two years, with a recent J.P. Morgan report estimating an increase of more than 400% since the start of 2024. Unfortunately for consumers, the surge isn’t limited to DRAM chips, as an acute manufacturing shortage in flash memory has also pushed up prices for SSDs, flash drives, memory cards, and external storage.

– Daniel Lemire (@lemire) August 5, 2026

According to industry reports, contract and retail prices for NAND flash have risen between 50% and 100% since the start of the crisis, triggering price increases across a broad spectrum of consumer electronics, including laptops, desktops, graphics cards, smartphones, and video game consoles. While brands such as HP, Dell, and Apple have raised prices for their laptops and MacBooks, Sony and Nintendo have also raised prices for the PS5 and Switch 2, respectively.

Ominously for consumers, South Korean semiconductor giant SK Hynix has predicted that the global memory crisis is far from over. According to CEO Kwak Noh-Jung, 2027 will be the worst year ever for the industry, with demand continuing to outstrip supply even beyond 2030.

However, whether these predictions materialize will depend in part on the health of the AI industry, amid growing fears that speculative investment in the sector is creating a bubble that could burst at any time.

This spring, my wife and I moved from the Denver suburbs into the Colorado Rockies. The high-altitude retreat presented a welcome reprieve from our otherwise tech-dominated lives. As remote workers, however, we faced certain challenges, and technology was the only reason our move was possible in the first place.

Prior to my time at El Reg, I, among other things, breathlessly covered a subset of the networking industry, which as it happens had become big business following the COVID-19 outbreak and the pandemic that followed. I’m speaking, of course, about software defined wide-area-networking (SD-WAN).

Little did I know, a mere six years later, this technology would become essential to my wife and me.

Had it not been for 5G and Starlink, the move might as well have been time travel. To put it into perspective just how far into the foothills we’d ventured, the fastest wired connection at our disposal was DSL. CenturyLink promised, but notably did not guarantee, up to 20 Mbps down and a staggering 1.5 Mbps up.

Compared to the symmetrical gigabit fiber line we’d enjoyed in the Denver suburbs for the past six years, relying on DSL alone felt positively medieval. To those, including many of my neighbors, living this reality, you have my sympathies.

Thankfully, DSL wasn’t our only option. For better or worse, 5G and Starlink are helping to close the digital divide in a meaningful way — but individually they’re still far from perfect. 

Where we live, T-Mobile’s 5G internet service delivers near-gigabit download speeds and uplink speeds ranging from 30–40 Mbps. Meanwhile, Starlink promised between 100 and 400 Mbps downlink depending on the plan, with uplinks that, at least for us, capped out at around 30 Mbps.

Either would be fast enough for our needs, but reliable? Turns out there’s still some room for improvement. In the past two months we’ve experienced no fewer than a dozen dropped connections lasting anywhere from a minute all the way to five. Making matters worse, the majority of these drops took place during working hours.

With a relatively clear view of the sky, we’ve found Starlink to be more reliable, but even SpaceX isn’t immune to the occasional outage. That’s not to mention the threat of space junk kicking off a Kessler syndrome-like event.

SD-WAN to the rescue

The easy way to approach multi-WAN would have been to set up two networks, each with its own SSID. In the event one went down, we’d switch to the other. 

While simple, it’s far from perfect, and does nothing to mitigate disruptions to the voice and video calls on which my wife and I spend a considerable amount of our week. With two WiFi access points, we’d also have to contend with interference from overlapping channels.

Wrangling multiple LANs, wireless SSIDs, and firewall rules isn’t exactly ideal. Thankfully, there’s a better way, and we just so happened to have all the equipment we needed.

SD-WAN was developed in the early 2010s to solve a very specific problem: as SaaS apps like Office 365, Salesforce, and Google Apps took off, more and more enterprise traffic was going out to the internet. 

This was a problem for a lot of enterprise WANs, which often relied on expensive MPLS links designed to connect satellite offices to servers running back at headquarters or regional datacenters. In many cases, these WANs were built to prioritize reliability over bandwidth, and due to their topology required backhauling traffic tens or hundreds of miles just to access the internet.

SD-WAN gateways offered an alternative. These devices could route internal traffic over the private WAN while SaaS and other internet-bound traffic could be piped over whatever local ISP served businesses in that area.

SD-WAN is an entire can of worms in itself, but for my purposes two key capabilities stood out: multi-WAN and policy-based routing support. Our router, a Ubiquiti UniFi Dream Machine SE (UDM-SE), already supported both.

The hardware

The UDM-SE is not Ubiquiti’s newest nor even its most capable gateway. Introduced in 2022, the all-in-one appliance combines a 3.5 Gbps router (with IPS/IDS enabled) with an 8-port gigabit PoE switch, an integrated network video recorder, and critically two WAN ports, one good for 2.5 Gbps and another capable of 10 Gbps. 

Ubiquiti’s SD-WAN implementation is somewhat basic compared to gateways from HPE, Cisco, or Extreme. However, of the bevy of networking and security features SD-WAN encompasses, we only really needed a handful – multi-WAN and policy-based routing were the main ones – and Ubiquiti is kind enough not to gate this functionality behind an enterprise license.

The setup itself was about as simple as it gets: plug each ISP-provided gateway into one of UDM-SE’s WAN ports and tell it whether you want to load-balance traffic across the two, or failover in the event of an outage or dropped connection.

In practice there were a few extra steps, including disabling the 5G and Starlink gateways’ onboard routing and WiFi functionality. Starlink supports a bypass mode, but the T-Mobile modem+router combo didn’t. We were able to disable the onboard WiFi, but only using an onboard utility.

Putting it to the test

With everything connected, we opted to configure the UDM-SE for WAN failover with T-Mobile as the primary and Starlink as the backup, rather than load-balancing traffic across the two, in part because the 5G link was so much faster.

The first few days after setting everything up, my wife and I were questioning whether we even needed Starlink. But then we saw the notification: “Internet connection WAN1 (T-Mobile USA) on port 9 is down and WAN2 (Starlink) is now active,” and then a few minutes later, “Internet connection WAN1 (T-Mobile USA) on port 9 is restored after failing over to WAN2 (Starlink).”

Both notifications were dated two hours earlier. We just hadn’t noticed. In fact, despite at least a dozen dropped connections over the past two months, the UDM-SE failed over fast enough that it was never an issue. No dropped calls, no buffering video, no timed-out web pages.

Perhaps the most surprising discovery was that while Starlink didn’t offer the fastest or lowest-latency connectivity, it was remarkably reliable. Despite less than perfect alignment and view of the sky, the logs show just two dropped connections in the past month, one of which happened in the middle of the night when no one was awake to notice.

While WAN failover kept us online, it didn’t do anything to guard against network congestion. The UDM-SE offers a preconfigured quality-of-service (QoS) policy that in theory should detect Zoom or MS Teams traffic and prioritize it over something like Windows Update or YouTube.

But since we had two WANs, we could go one better and take advantage of another SD-WAN favorite: policy-based routing. This allowed us to force certain devices or services to prioritize one ISP over the other. For example, because Starlink had proven to be the more reliable of the two links, I wrote a policy to route traffic from my wife’s work laptop over the satellite link first and fail over to T-Mobile in the event of an outage. This also had the benefit of ensuring that if I happened to kick off a large game or AI model download while she was in the middle of a meeting, I wouldn’t get myself in trouble for turning her Zoom call into a pixelated mess.

The digital divide is narrowing, not closed

In the more rural parts of the US where wired connectivity isn’t a given, cellular home internet and low-Earth orbit (LEO) satellite communications, like Starlink, have gone a long way toward closing the digital divide — particularly for those working from home.

However, individually they’re still far from perfect. Dropped connections weren’t a daily occurrence, but they still occurred more frequently and for longer than we would have liked. Five minutes is a long time to wait wondering whether the person on the other end of the Zoom call will be there when you eventually get back online.

Technologies like SD-WAN can help to mitigate the problem by allowing multi-WAN failover, load balancing, and policy based routing, but to take advantage of them also requires paying for two ISPs and hundreds of dollars of equipment, which may or may not require monthly subscriptions or annual licenses.

We were fortunate that the hardware we already owned supported these features and didn’t lock them behind a pricy enterprise license. Even still, we’re paying more for internet than we’d like. Between T-Mobile and Starlink, we’re spending around $125 a month to get online.

Along with the higher price, network latency is also a lot higher than we’d like. Compared to our old fiber line, which routinely achieved latencies under 5 ms, T-Mobile and Starlink ranged from 15 ms at best to as much as 80 ms at worst.

The only online game I play regularly is a real-time strategy that’s old enough to drink. I’m not playing CS:GO or whatever twitch shooter is popular these days, so the higher latency wasn’t that big of a deal, but if you’re a competitive gamer, a wired connection is still a must.

The good news is T-Mobile and Starlink may not be our only options for high-speed internet for much longer. There are numerous ongoing efforts across the US and other nations to improve access to fiber connectivity, including one slowly working its way down the highway in our general direction.

On a personal note, while it’s neat to experience a technology firsthand that I dedicated years of my life to writing about, the moment fiber-to-the-home comes, both our Starlink and 5G home internet subscriptions are getting the boot. ®