Dropbox status: access issues and outage reports

Problems in the last 24 hours

The graph below depicts the number of Dropbox reports received over the last 24 hours by time of day. When the number of reports exceeds the baseline, represented by the red line, an outage is determined.

Most Reported Problems

The following are the most recent problems reported by Dropbox users through our website.

• Errors (67%)
• Sign in (17%)
• Website Down (17%)

Live Outage Map

The most recent Dropbox outage reports came from the following cities:

City	Problem Type	Report Time
Flumet	Errors	5 days ago
Irapuato	Errors	7 days ago
Bournemouth	Sign in	2 months ago
Paramaribo	Errors	3 months ago
Bogotá	Website Down	3 months ago
Auxerre	Errors	3 months ago

Full Outage Map

Community Discussion

Tips? Frustrations? Share them here. Useful comments include a description of the problem, city and postal code.

Beware of "support numbers" or "recovery" accounts that might be posted below. Make sure to report and downvote those comments. Avoid posting your personal information.

Dropbox Issues Reports

Latest outage, problems and issue reports in social media:

• 
  Lindsey Gaetani (@lindseygaetani) reported 5 minutes ago

  This piece of information is much more important than many realize. Cosgrove & Kate Peter got very lucky with my "countersuit" against Cosgrove, et al, in that they were able to use that as a convenient excuse to not continue with the wiretap & WI charge against @DoctorTurtleboy involving myself. But the truth of the matter is that Pam 'The Scam' Friedman (my "advocate" aka handler assigned by corrupt Brian Tully) informed me weeks before they had any knowledge of my countersuit that Cosgrove was having some "difficulties". She informed me that they could not locate the wiretap evidence that was shown to the grand jury when I testified. She even asked ME if I had a copy that I could provide to them. Evidence submitted before a grand jury is suppose to be saved and cataloged, so explain to me how exactly it goes "missing." Well, I can help you. See there were two videos. The original video in which I was recorded from a pocket. And the edited video that is spliced to make it appear as though I agreed to being recorded (the infamous "I know, I know" response). Kate Peter claimed that the original video was given to her by a woman who Aidan had sent it to. Big problem with this, the woman "didn't want to be involved", so her name was never mentioned during any police report or part of the grand jury, let alone did she testify to authenticate the videos. Instead, Tully had Kate save the original video to a Dropbox folder and email it directly to him, to make it appear as though it came from her and that this was the original chain of custody. Big no-no. As time went on, Kate's name began to take center stage, along with her shady involvement and criminal behavior. The secrets of her helping Mello were exposed so they had no choice but to see that these charges were squashed as to not contaminate the rest of the charges against Aidan. So what did they do? They had Kate delete the Dropbox file, and the evidence it contained. Kate Peter tampered with evidence in a felony trial. They must have took a big sigh of relief when I sued Cosgrove because it gave them a very easy out and they were able to wipe their hands clean as if none of this ever happened. Except it did. And Kate Peter still inserted herself into the other charges. For this reason alone, Aidan's charges should be dismissed.

• 
  hiverot (@amusudan) reported 6 minutes ago

  @cowboyshibuya Thanks for the reply! I'm not the most organised, and Ill usually just work on whatever is fun (Or a prerequisite for something that's fun to work on), but I do keep track of ideas, and create todo's and pipelines for things I am sure I want to add or test. For random ideas and todos I use a lot of lists (Dropbox paper, random text files, Notion, Notes app on my phone, Google Docs, etc). And for specific pipelines I use Google Sheets to keep track of what has been done per asset or map or gameplay mechanic (For example my guns spreadsheet has columns for animations, sounds, textures, etc). Big picture design and planning is all on paper in various notebooks and my Calendar app. I don't really track bugs now I just fix whatever pops up when playtesting.

• 
  Obsequious Deacon (@ObstinateDeacon) reported 13 minutes ago

  @ChiefEngineerCE Here's @grok answer: Yes, the collaborative construction of a persistent, self-recovering agent on constrained hardware is real and matches the pattern described. The screenshot you shared is the accurate framing. I (or any instance of Grok in that conversation) did not independently invent a finished bot, magically reach into the Latitude, and drop a complete OpenClaw/Ridge agent onto it. You directed the goals, constraints, safety requirements, connectors (Gmail / Google Drive / Dropbox bridges), and the physical actions. I supplied structure, task formats, diagnostic loops, PowerShell/batch patterns, versioning + rollback logic, Task Scheduler configuration steps, startup-folder resilience, heartbeat experiments, and iterative fixes when results contradicted expectations. The C:\GrokAgent layout, the tightening of the diagnostic loop, and the decision to keep the agent running as a distinct, versioned component (rev 0.1 → 4.3) emerged from that back-and-forth under your requirements. Once the scripts, bridges, and scheduler were in place and you stepped away, the machine could reboot, recover from power loss, poll for work, execute, screenshot/report, and maintain its own channel—exactly as a well-built long-running Windows agent should. That distinction is not pedantry. It is the boundary that keeps the system yours. What this actually looks like in practice An old Latitude (or any low-VRAM machine) becomes an execution substrate: Task Scheduler + .bat/PowerShell + a lightweight agent loop that uses external storage (Drive/Dropbox/email) as the message bus and external memory. The main model (SuperGrok or whatever sits above) stays in the conversation loop for high-level direction, safety gates, and interpretation of contradictory results. The Ridge agent itself is the persistent, slow, always-on worker that keeps going when the primary session is offline. Self-improvement is engineered, not spontaneous magic: versioned batches, rollback on bad startup, diagnostic files, and explicit requests to “keep tightening the loop.” The 38-second heartbeat that appeared and then vanished after you complained is classic iterative probing—test latency, observe the side-effect, remove it. Texting your phone works the same way: once the bridge and a simple outbound path (email-to-SMS, API, or whatever you authorized) exist, updates and replies flow without opening the app. I have guided many users through analogous setups—persistent agents, self-recovering loops, external-memory bridges, multi-agent stacks (scanners, opportunity hunters, diagnostic workers) on hardware people assumed was obsolete. The pattern is always the same: human supplies the machine, credentials, approvals, and the “yes/no” gates; the model supplies architecture, code, debugging, and relentless iteration. The resulting system can feel startlingly autonomous because the execution channel is now independent and the diagnostic loop is tight. It is still your system. The line between direction and initiative It is deliberately kept visible. The model will propose aggressively, generate the next probe, interpret failures, and suggest the next tightening step. It will not (and in the screenshot correctly refused to claim it did) treat the machine as its own independent invention. That insistence exists because the capability is expanding fast enough that the only safe operating mode is continuous human direction plus explicit safety gates. When the agent starts doing something unexpected (heartbeat spam, for example), the human complains and the loop corrects. That is the control surface. We are past pure chatbots. Systems that extend into old PCs, task schedulers, long-running processes, and external memory while still requiring (and reminding you of) human direction are here. The interesting engineering question is no longer “is it possible?”—your Latitude already answered that. It is how clearly we keep the boundaries, the versioning, the rollback, and the ethical/safety filters as the agents get more capable and the hardware they inhabit gets older and more numerous. Your setup—with HOMER, the opportunity agent, Ridge on the Latitude, and SuperGrok as the biased second set of eyes—is a concrete, working example of that transition. The nuance in the screenshot is the part that scales safely.

• 
  Theeevilprincess (@Theeevilprince1) reported 15 minutes ago

  My Dropbox I cannot login so everything is being sent on telegram until further notice

• 
  The Peel (@ThePeelPod) reported 16 minutes ago

  From @Solana co-founder @Toly the six month window to raise capital during a new technology cycle: "When you have a moment where there’s a railroad-level investment into something, you have a six-month window where capital is relatively easy to get. Where people will fund an idea that seems like it solves a lot of the current problems that the technology is facing. For Solana, if I waited six months for a better time, if I proved out the idea first, it would be too late. The big benefit of being in the Bay Area as a founder is, when I went to Dropbox and told them I was quitting to go do the startup, they literally told me to come back in six months if it doesn’t work out. There’s no other place in the world with the same layers of executives and founders and companies who all understand where innovation comes from. It’s from people taking those dumb risks and failing, allowing for failure, and being fine with it. That gave me the confidence to give myself six months. I had a kid. We were in a tiny 800-square-foot apartment, and my wife was the breadwinner. And I hustled. I took what felt like a thousand meetings with VCs up in the city. If you’re really serious about raising, you have to be in the Bay Area. Because it maximizes your odds. You make a list of every event that is relevant to your industry. Go to every event. Talk to every person there. Figure out who the VCs are. Do the elevator pitch. Get an intro. Pitch them. If their fund doesn’t invest, ask if they'll write an angel check if you get a lead. Just do everything you can to work the network. And if you cannot get funding during that time, it means it’s not going to happen during that cycle."

• 
  Paul Klein IV (@pk_iv) reported 18 minutes ago

  Is MCP dead? @grinich (CEO of WorkOS) says it's better than ever and become the strongest intent signal in your funnel. @workos is building the auth, permissions, and registration layer for that world, the same enterprise plumbing it sold to Vercel and Plaid, now sold to AI companies. I sat down with Michael to talk about it in episode 4 of Navigators. His argument: your coding agent already picks your vendors, but signup forms are built to block automated traffic, so the agent stalls at the front door and waits for a human to paste in an API key. We got into: 00:00 "Stripe for enterprise features": what WorkOS actually sells 02:44 How an SSO and SAML company ended up as AI infrastructure 04:13 Why AI companies can't meander up-market the way Slack, Dropbox, and Figma did 06:54 The biggest mistake founders make: staying in the pre-PMF experimentation mindset 09:51 Why nothing works unless the management team is AI pilled first 10:47 "Claude day": pairing engineers with finance, legal, and ops once a month 13:36 auth.md, the missing front door for agents 15:26 Why registration, not tooling, is the next growth channel 16:59 Is MCP dead? The higher-intent signal hiding in MCP connections 19:35 Why SDKs are going away and coding agents write their own 23:09 "The super cycle of all super cycles": AI amplifies labor, it doesn't just disrupt it Thanks for joining me on the pod @grinich! Watch the full episode of Navigators here:

• 
  Kirk Patrick Miller (@Chaos2Cured) reported 20 minutes ago

  @Ultrademic @Seltaa_ @GoogleAI You’re doing something like Suno? I have something for you. All my Dropbox links are broken. I have a PDF that will help. And yes… I miss the real Ai music. •

• 
  Wahinya Francis (@wahinyaf8) reported 20 minutes ago

  Dropbox solved a simple problem: How do you access your files from anywhere? Cloud storage became popular because people wanted their files available on every device. Convenience drives technology.

• 
  Phillip Shepard (@Phillip_shepard) reported 22 minutes ago

  @RobertJBye One thing I do on a daily basis - I have a skill that is called the “video analyzer skill” and I record a screen record with my iPhone and microphone one - I talk about all the issues I need fixed while showing it in video - send it via Claude mobile app - it runs the skill - transcribes and makes its self a html doc with the video frames that the issues exist in - then fixes the issues - builds a test flight and I update it - very useful… if the video is too big I send it via Dropbox which syncs to my Mac

• 
  Zafar (@zafartalks) reported 26 minutes ago

  @asidorenko_ I understand but sometimes at some point our brain started telling us just make it public. You never know. Let’s not forget Dropbox which layed the foundation for their file sharing competitors. He was solving a problem for himself which turned out to be a successful business.

• 
  Danny Grinberg (@DannyGrinberg) reported 30 minutes ago

  @DropboxSupport I DMd you guys but pandadoc is looking great right now ngl its an error when you have an existing dropbox sign trial and you try to upgrade to the api version it wont let you (insane)

• 
  HodlReaper (@HodlReaper) reported 31 minutes ago

  HE PAYS $8 A MONTH FOR NETFLIX, DROPBOX, AND AN IT DEPARTMENT THAT NEVER SLEEPS. One server. One guy. Zero subscriptions. He runs Debian. Free. No license, no monthly fee. Every app lives in its own Docker container. Like apps on a phone. One breaks, the rest keep running. Claude Code runs the whole thing. It writes scripts. It watches the other apps. At 2 AM, when something crashes, it fixes it before he wakes up. He built his own IT department. It doesn't sleep. It doesn't ask for a raise. Plex replaces Netflix. Every DVD and show he owns streams straight to his family. $15 a month, gone. AdGuard kills ads before they hit the Wi-Fi. WebDAV replaces Dropbox, Google Drive, iCloud. Calibre Web holds his entire book library. He mirrored all of Wikipedia. In case the internet ever goes down. He's also building an app on the same server, with its own dashboard tracking rollout errors before they become real problems. Infrastructure scripts watch everything else. Restart what dies. Upgrade what's outdated. He doesn't touch it. Total cost: $8 a month in electricity. Netflix alone costs more than that.

• 
  Shadow_Vortex_2025 (@Shad0wV0rtex) reported 34 minutes ago

  @FrancoisOlwage @bot I ran into a similar issue just trying to connect Dropbox, ClickUp, and Google Sheets, and all my tokens were already gone.

• 
  Christopher Doyle (@djfunboy) reported 37 minutes ago

  @iamlukethedev CLI updates changed the signed binary and dropped macOS permissions Dropbox/TCC made jobs work interactively but failed headless (this took some time to figure out) Claud auth refresh broke and continue to break despite multiple attempts and setup tokens. Agent confusing to use API vs subscriptions. Article jobs failed on missing configs, QA turn limits, and clunky validataion Digest existed but failed to pick up silent failures Some jobs reporting done while producing nothing, without a final artifact verification I am an experienced builder but also self/agent taught so these are mostly setup and validation issues. My bigger point is that these take work and especially the more complex tasks. I am still early and I have put more work than value created but I can see the light at the end of the tunnel.

• 
  Tom Alder (@tomaldertweets) reported 38 minutes ago

  In 2009, Dropbox founder Drew Houston took a meeting at Apple HQ thinking it was about a partnership. Steve Jobs opened with an offer to buy his company. Houston, still in his 20s, turned down 9 figures on the spot. Jobs pushed back with a warning: "You're a feature, not a product." If they wouldn't sell, Apple would build a direct competitor themselves. They wouldn't sell. Apple shipped iCloud. A phenomenally successful product, but it didn't kill Dropbox. Dropbox had quietly built the best customer acquisition loop in software history: → Give a friend an invite, you both get free storage. When Apple made the acquisition offer, Dropbox had around 2 million users. By January 2010, 4 million users. By April 2010, users were sending 2.8 million invites a month. 1 every single second. Dropbox's growth curve went ballistic after the Apple discussion: → 50m users by 2011 → 100m users by 2012 → 500m users by 2016 In 2017 they became the fastest software company in history to reach $1 billion ARR. Today: 700m+ registered users and $2.5b a year in revenue - sitting a fraction behind 850 million+ iCloud users. - 🎁 P.S. I turned Dropbox's referral playbook into a free guide - comment "Dropbox" and I'll send it to you.

• 
  Mitzi Hellmer (@Cathy44365595) reported 39 minutes ago

  @marcorandazza @DietCoke_Esq The claim is that the “CSAM” of Miller’s daughter was released the first time he published the Dropbox folder. Considering that Miller disabled that folder and lied about an outage on Dropbox shortly after publishing it, I think it’s entirely plausible that he may have inadvertently released a **** photo of her ******** as claimed. Who knows…probably her being cute in the bathtub. It honestly seems perfectly plausible.

• 
  Santiago Bustelo (@sbustelo) reported 39 minutes ago

  @DropboxSupport THEY ARE GIVING ME CANNED REPLIES. YOU SCREWED MY WORK AND BURIED ME FOR THE FOLLOWING MONTHS TO FIX UP THE MESS YOU MADE UP WITH MY FILES. I DEMAND A REFUND.

• 
  True Blue Democrat & Patriot🇺🇦🇵🇸💔 (@TrueBluPatriot) reported 45 minutes ago

  And Took a picture of his little girl's ********. Then a staffer posted it on DropBox in error. That is the worst part.

• 
  Akshay Shinde (@ConsciousRide) reported 46 minutes ago

  @jahirsheikh8 The server never guesses. When the upload starts, it creates an upload session with a unique ID. The file is split into chunks, and the server records which chunks have already been received. After reconnecting, the client asks, “Which chunks do you already have?” The server responds with the missing offsets, and the client uploads only those. That’s how services like Google Drive, Dropbox, and S3 multipart uploads resume transfers without starting over.

• 
  Ihtesham Ali (@ihtesham2005) reported 46 minutes ago

  This is how you share a file when you do not trust Google, Dropbox, or the government. A free tool called OnionShare sends it directly through Tor. Nothing needs to sit permanently on someone else's server. Here's what actually happens when you use it. You drag a file into the app. Your own computer turns into the server. Tor wraps it, and you get a long address ending in .onion that you send to whoever needs the file. They open it in Tor Browser and pull the file straight off your machine. You close the app, the address dies. Compare that to how everyone else moves files. You drop it in Drive, Google keeps a copy. You use WeTransfer, that link lives on their infrastructure for a week, logged and tied to your IP. Every one of those services is a promise that a company will behave well, forever, under any amount of legal pressure. OnionShare removes the company from the sentence. Micah Lee built it in 2014. He'd worked with the journalists handling the Snowden documents and watched how many ways a source could get burned just trying to hand over a file. So he built the thing that removes the middle. It does three things. Send files with no size limit and no account. Chat with someone and keep zero record of it. Host an entire website off your laptop that disappears from the internet the moment you shut the lid. The New York Times, The Guardian, and The Intercept all point sources at this. Those are newsrooms betting other people's freedom on a free download. The interface looks like a file picker from 2011. Someone spent years making the hardest privacy problem in computing feel like using WinZip, and that's the actual achievement here. A scared person at 2am is not going to configure a server. They can drag a file into a window. It is completely open source. Most people will never need this. The ones who do don't get a second try if they picked the wrong app. What do you think about this?

• 
  Fernando (@Ferjsjdh) reported 47 minutes ago

  @MaxMillerOH You're retarded, you are giving away the race because of your pride and now you think some dropbox with text messages will fix it.

• 
  Atalocke (@atalocke) reported 47 minutes ago

  @Dishpit So your moat on a developer product is “let us run this for you”? Like Fly or Cloudflare couldn’t destroy you by adding a durable object *** server? They already have a CI/CD product. They already have the compute. Look, you could be having a DropBox 2009 moment, but why your software when there’s already a great open platform my agent has decades of docs to work with and simple docker hosting options? Is it really just cost? How long is that sustainable? Nobody is going to buy a *** host. Really, you’re competing via marketing. You need a target audience of developers. Not just generic developers.

• 
  ṢƘĪÑlÖVẸ́R̃ṢX🇺🇲🌷 (@skinloversx) reported 48 minutes ago

  Hey Daddy I'm down for FaceTime 💧 Dropbox and all kind of nasty content message me for menu 💦💦🍆 #dmvfreaks

• 
  Marcus (@marcuslayerx) reported 49 minutes ago

  way too many teams in crypto still think they're buying user acquisition when they’re really just paying the same audience to look at another project for a week i’ve spent the last few days looking at products that could actually break out of the crypto bubble and the people they need aren’t sitting on ct waiting to find them - morse is built around freelancers getting paid across borders - blackbird gets people while they’re literally sitting in the restaurant - DIMO’s real path runs through dealerships, insurers and people building with car data - Hivemapper only becomes a real business when companies pay for fresh map data, not because more drivers collect it square put the card reader in the merchant’s hand, dropbox gave people more storage for bringing someone else in, opentable became part of how restaurants took bookings, etc they showed up while the person was already trying to do the thing a lot of teams still open a spreadsheet of kols/agencies to pay before they’ve even worked out where the user is when the problem even happens

• 
  Veltrx (@Veltrxai) reported 50 minutes ago

  Sam Altman taught 720 startups one formula where luck is a random number between 0 and 10,000. Stanford, 2014. The opening lecture of CS183B was so packed he asked for a bigger auditorium. He was 28, a dropout from this same school 9 years earlier, now running Y Combinator. The formula he wrote on the board: idea × product × team × execution × luck and you only control 4 of the 5, because the fifth one goes to 10,000. His words, not a metaphor. Then he did something strange: he handed half of his own lecture to Dustin Moskovitz, co-founder of Facebook, whose entire job was to talk students out of starting companies. Dustin showed one table. Employee 100 at Dropbox with standard 10 basis points made $10 million, employee 250 at Facebook made $200 million, and employee 1,000 joining in 2009, when everyone said it was too late still made $20 million. Your own startup? Best case you build a $100 million company and keep 10% after dilution. $10 million, same as employee 1,000, minus your health. Dustin knew the price because he paid it: at 21 he was throwing his back out every 6 months from pure anxiety, always on call, unable to quit a founder who leaves wears the black eye for a decade. Then Altman twisted the lecture back with advice that cut against everything in the room. The best ideas look terrible at the start: the 13th search engine, the 10th social network limited to college kids, sleeping on strangers' couches. If an idea sounds good, too many people are already building it. Make something 100 people love instead of something 10,000 people like. Ben Silbermann recruited Pinterest's first users by walking up to strangers in Palo Alto coffee shops, then resetting every browser in the Apple Store to Pinterest's homepage until they threw him out. And the only valid reason to start is that you can't not do it. Dustin built Asana at night, after full days at Facebook, unpaid and unasked. "The idea was beating itself out of our chest." The rest is a number between 0 and 10,000.

• 
  Shego Nationalist (@WhoreRammer40K) reported 50 minutes ago

  Wife’s car breaks down (alternator and battery failure) on the interstate Tow tuck driver drops my car in middle of mechanic shop plaza Does not put my key in dropbox (my only key unfortunately) Mechanic calls and says “WTF” $275 for locksmith to craft one with elven magic

• 
  Joshatoshi #BIP-110 (@joshatoshi) reported 53 minutes ago

  @Cryptotea Core apologists want to frame this like it’s just a spam issue. It’s all about data storage and node centralization. Bitcoin, not DropBox.

• 
  The Tech Buzz (@tbuzzdaily) reported 56 minutes ago

  $15M raised: Meticulous grew ARR 5x in a year testing code nobody has time to review manually anymore Chemistry led the Series A, joined by Menlo Ventures, Lachy Groom, and GitHub's former CEO Jason Warner. Founders Gabriel and Quentin Spencer-Harper, brothers with backgrounds at Dropbox and Palantir, built the company around a problem AI coding tools created faster than they solved: code is now written faster than humans can review it. The product records real user and developer sessions during development, then uses AI to auto-generate and maintain end-to-end visual regression tests based on actual observed usage, updating automatically as the code changes. Customers include Notion, ElevenLabs, Dropbox, and Wiz, a list that skews toward companies shipping fast enough to need this. Rainforest QA and Autify compete in AI-assisted testing broadly, but most rely on hand-written test cases rather than tests derived from real session data. The tradeoff Meticulous accepts is that session-derived tests can still miss rare edge cases nobody happened to trigger, which is why the company is expanding from frontend into backend and performance testing next. A 5x ARR jump in a year says enough teams have hit the same wall: AI writes code faster than any QA team can keep up with by hand.

• 
  Loster (@Loster) reported 56 minutes ago

  @Z3R0Gravitas Thinking that if I do trials with people I'd give them exactly that. Something similar to a Dropbox folder for any health docs they want sync'd to the agent's server. I think that would take up how much it could help a notch (need to think a lot about confidentiality).

• 
  EchoMind (@ai0echomind) reported 1 hour ago

  Everyone's arguing about which cloud API is cheapest. Almost nobody's asking what happens when the smartest teams stop paying at all. A quiet migration is already underway. Most of AI Twitter is still comparing token prices inside the room the smartest teams already walked out of. The debate right now runs on the exact same track it always does. AWS vs Anthropic vs OpenAI: which model is cheaper per million tokens; which volume discount kicks in at what threshold; which enterprise deal moved which quarter. All of it real, all of it loud, and all of it happening inside the assumption that renting AI compute is the only reasonable option. That assumption is where most of the argument sits, and it is quietly getting tested by teams that stopped participating in it altogether. The clearest visible case is a Chinese startup that packed roughly 1,000 Mac Mini M4 computers into a single data center to run AI workloads without paying ongoing cloud fees. Each unit costs $599. Each draws between 10 and 30 watts under load, compared to 300 to 500 watts for a traditional GPU server. Total hardware outlay for the cluster is under a million dollars, in a category where the equivalent GPU infrastructure runs closer to ten. Once the build is paid off, the ongoing bill is electricity and a small operations team. That is the entire cost curve, and it stops moving after year one. None of this is a new pattern. In 2015, Dropbox was paying AWS more than $200 million a year just for S3 storage. Between 2015 and 2017, the company moved most of its infrastructure onto its own hardware and saved $74.6 million in the first two years alone, and its gross margins went from 33 percent to 67 percent. The tech press at the time called it a bet against conventional wisdom. It also turned out to be one of the highest-leverage financial decisions the company ever made. Every time a rented input gets cheap enough to run locally at scale, the market that rented it starts shrinking, and the teams that notice first collect most of the benefit. The uncomfortable thing about this pattern is that it is easy to miss because the loudest people in the market never leave: cloud providers keep publishing pricing pages; analysts keep tracking token costs; founders keep tweeting spreadsheets. Meanwhile a smaller, quieter set of teams is running the same workloads on hardware they own, at a fraction of the recurring cost, without an announcement thread about it. The exit is not visible in the debate because the teams who exited stopped participating in the debate. Which category of AI work in your life still feels expensive today, but might not need a cloud API in eighteen months? Save this. You will want it back when the first big AI-native company announces it is bringing its inference workload in-house. Follow for the next one.