Amazon Web Services status: access issues and outage reports

Problems in the last 24 hours

The graph below depicts the number of Amazon Web Services 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 Amazon Web Services users through our website.

• Website Down (71%)
• Sign in (14%)
• Errors (14%)

Live Outage Map

The most recent Amazon Web Services outage reports came from the following cities:

City	Problem Type	Report Time
Boca da Mata	Errors	3 days ago
Township of Evan	Website Down	17 days ago
New York City	Website Down	20 days ago
Ciudad Jardín	Website Down	2 months ago
Kyiv	Sign in	3 months ago
Chennai	Website Down	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.

Amazon Web Services Issues Reports

Latest outage, problems and issue reports in social media:

• 
  CaptainAmericaTex (@CaptAmericaTx) reported 8 minutes ago

  @LayoffAI During the last 12 months Amazon AWS had outages directly caused by initiatives from the AI group (with associated revenue impact). No surprise they are down sizing the snake oil sales men group.

• 
  Priyanshu (@iproductAI) reported 9 minutes ago

  I requested @awscloud to increase my Opus 4.6 V1 limit and I’ve been chatting with the AWS support team for almost 4-5 days now. They’re telling me this. Is @awscloud a government company? I mean, you guys can’t just pass the problem from one department to another. I mean, WTF? Now I have to raise my query again to sales? Why can’t you just pass this query? You already have more context about what the issue is. I can’t believe how these big MNCs are working these days. Totally absurd service from @awscloud. One more thing, please educate your support. I mean, she didn’t even know what the TPD limit is in the service quota. She literally replied the first time saying there’s only a TPM limit and no separate TPD limit.

• 
  Prateek Gupta (@p_valuee) reported 9 minutes ago

  @AWSSupport My entire production is down since 3 days, I believe this is a P1 and should be treated like one @AWSSupport. Please help me with an ETA

• 
  Ashmit Dutta (@ashmit105) reported 10 minutes ago

  @awscloud keeps charging me $60 a month. I can’t login to my account due to a deprecated email. Making support tickets goes nowhere. Anyone got advice on what to do?

• 
  GettingMyGlitchOff (@GettingMyGlitch) reported 11 minutes ago

  @AWSSupport Again there is a mass wave of suspensions like in May, I'm assuming in error as I've been on this platform for over 10 years and this seems to be happening more frequently. Emails from mturk-noreply havent worked in a long time so any updates on this situation?

• 
  Keith (@happy_keith) reported 12 minutes ago

  @amazon @awscloud @AmazonUK What does it take to get you to respond to a serious problem DO NOT tell me to go to customers services AI as it is incapable of responding to the issue This is a complete farce !!!!!!

• 
  Steven Coburn (@pcgamer11) reported 13 minutes ago

  @PlayStation We can't play the open beta rn because PSN IS DOWN. Yes, it uses/runs on @awscloud but you could at least ACKNOWLEDGE the issue!!!

• 
  Abhijit Tripathy (@AbhijitTripat13) reported 13 minutes ago

  @awscloud @awscloud still no response. You guys are so slow

• 
  Andrii Bidochko 🦉 (@andriibidochko) reported 19 minutes ago

  The Missing Primitive for Autonomous AI: Bounded Agent Payments 💳🤖 For the past two years, autonomous agent loops (like @OpenClaw, Hermes, or custom agent harnesses) have been missing a critical infrastructure piece: the ability to transact on demand without a human in the loop. When an agent hits a paywalled research endpoint, a paid search API, or an MCP server mid-execution, the entire loop freezes. Up until now, your options were: 1. Hardcode expensive subscription API keys upfront. 2. Freeze execution and wait for a human to manually pay. @awscloud and the OpenClaw Foundation just published a blueprint solving this via Amazon Bedrock AgentCore Payments and the ⁠#x402⁠ protocol. Here is how it works, why it matters, and the primary architecture use cases:

• 
  Nilesh (@nile3h) reported 19 minutes ago

  Reached out to @awscloud team yesterday related to billing issues wrt AWS partners, no replies Can anyone help?

• 
  Jeremiah "SMEEgle" Castillo (@jcastillo4tx) reported 24 minutes ago

  @AWSSupport Live production server is down. Case 178699777500302 (Account Reinstatement) filed with a callback requested — call failed to initiate. Earlier case 178699433100264 still unassigned. Need urgent help, can someone DM me?

• 
  absk (@abskwdkr) reported 28 minutes ago

  My ec2 ssh logs in too slow and lags @AWSSupport

• 
  Arian van Putten (@ProgrammerDude) reported 31 minutes ago

  @QuinnyPig @awscloud Are they gonna fix people needing to know what region their orgs IAM identity center is deployed in order to set up cli login? (How the heck would an employee know this??)

• 
  matyas.dev (@DevMatyas28516) reported 32 minutes ago

  @AWSSupport Still nothing happened after a week, after I sent your X account our issues numbers.

• 
  SwarmOne.ai (@swarmoneai) reported 33 minutes ago

  OpenAI just unveiled Jalapeño - their first custom inference chip with Broadcom. "Substantially better performance per watt." Interesting. But a chip doesn't fix a misconfigured serving stack. Custom silicon with default batching, naive routing, and untuned KV-cache management is still burning money. SwarmOptimizer doesn't care what chip you run. It tunes the deployment layer - batching, caching, routing, concurrency - where most of the waste actually lives. The silicon is one layer. The optimization is another. Now on @awscloud

• 
  Open_ERV (@open_erv) reported 36 minutes ago

  Unfortunately although they appear to be awesome people BrambleCFD is turning out to not be that hot. The main problem is the relationships/what they do of all the different settings is ridiculously opaque. There is no documentation. Their solution is to try to explain things in a video call, and if you need help, you ask for it, which it takes a week or more to get any kind of answer from an actual human, not because they are doing anything wrong but that's just not a good system. It's a long long way from the useability of simscale. I did however uncover an option that might be reasonably good, which is a virtual machine that I pay for the core-hours on. In many ways this is better. I can work directly with openFOAM and one of the front ends on a high powered linux computer with hundreds of gigs of ram and 96 high powered cores, and still only pay for what I use, theoretically. The openFOAM foundation has a system worked out ad directly offers the service, unfortunately they in turn use the amazon AWS or the microsoft Azure system, but what can you do. There are other companies that do similar things, but they probably aren't as well done as the one from the actual foundation. I think I'll try that one first. Having an AI in a harness on the machine is probably going to be indispensable, but I'll be using it primarily as a learning tool rather than asking it to do everything for me. I have been able to set up CowAgent, which is kind of basic but seems to work ok, with DeepSeek. A "harness" allows the AI to run commands on your computer and read the output automatically, as well as the other things web chat stuff can do. Secondly, it can store information in files and run the AI in a loop, doing many inferences one after the other, thus getting far more done than a web chat can (actually they might do something similar now IDK but they don't seem to).

• 
  jd42 (@jdonovan42) reported 39 minutes ago

  @DarioCpx Pls recalculate considering the massive clouds the AI revs sit on. Each $1 in AI attaches $1.5-2.5X attached cloud biz and your down to 30-45% of revs. Then as open wgt models gain apply the 100% revenue retention vs. 60% on frontier and look at EBITDA % from frontier is 10-15% max. But nice try ;) No doubt Anthropic and OpenAI both stimulate demand for Amazon AWS. It creates 2x the cloud biz than it does the direct AI biz. So why not count the full picture of things vs just the #'s that fit one narrative.

• 
  Milan Jovanović (@mjovanovictech) reported 42 minutes ago

  @mrwcjoughin @awscloud @Azure I doubt this is a compute issue, looks like a bottleneck from past choices coming to bite you

• 
  Payal (@payal_codes) reported 42 minutes ago

  Day 1 : "How to Scale an App to 10 Million Users on AWS" If I have to design a system for 10 million users, I won't build everything on Day 1 because it will add unnecessary complexity and cost. I'll start simple with one application server and one database. As traffic grows, if the server starts reaching its CPU, memory, or storage limits, I'll first scale vertically by moving to a bigger instance. Once that is not enough, I'll separate the backend and database so both can scale independently. To avoid a single point of failure, I'll deploy the application across multiple Availability Zones and put a Load Balancer in front so if one server or AZ goes down, traffic is automatically routed to healthy servers. As the number of users keeps increasing, I'll make my application stateless by storing sessions in Redis. This allows me to add multiple application servers behind the Load Balancer and scale horizontally. If my database starts getting overloaded with reads, I'll use Redis to cache frequently accessed data and add read replicas to distribute read traffic. For static assets like images, CSS, and JavaScript, I'll store them in Amazon S3 and serve them through CloudFront so requests don't keep hitting my application servers. If traffic suddenly spikes during sales or events, I'll enable Auto Scaling with CloudWatch metrics so AWS automatically adds or removes servers based on demand. As the application becomes larger, I'll split the monolith into microservices. This allows each service, like authentication, payments, or notifications, to scale independently instead of scaling the entire application. If the database becomes the bottleneck, especially for write operations, I'll use sharding or federation depending on the data and business requirements. Finally, when users are spread across the world, I'll deploy the application in multiple AWS Regions to reduce latency and improve availability. My approach is always the same: find the bottleneck, solve that bottleneck, and only introduce more complexity when the current architecture can no longer handle the traffic.

• 
  mail box (@mailbox28564784) reported 44 minutes ago

  @amazonIN @amazon @awscloud I am extremely disappointed with Amazon's customer service. Despite returning the product, I have neither received my ₹1,400 refund nor a replacement. Repeated follow-ups and social media complaints have not resolved my issue.l

• 
  GamerUP (@GamerUPGaming) reported 46 minutes ago

  @TheRavenHelm man... are you detective seeds right? this is a new profile name ? whatever.. the "digital only direction" was announced more than 3 weeks ago.. not last week.. PSN servers are hosted by Amazon AWS, and amazon AWS had some disruptions today, they already solved the issue.

• 
  Mr Sage (@GreatSage_0x) reported 56 minutes ago

  Custos makes verdicts on AI agent transactions. Problem: it had no memory. Every decision started from zero. Meet Anamnesis — I gave it a memory layer using CockroachDB. Built for the @CockroachDB x @awscloud Hackathon

• 
  metis (@metis00001) reported 56 minutes ago

  I can't understand why AWS has not been able to resolve my issue of payment method. Whatever credit card I put, they are not able to verify it and they keep asking me to call the card issuer. It's not card issuer issue! @awscloud

• 
  Synapse Brief (@Synapse_Brief) reported 57 minutes ago

  @yugacohler @awscloud @CoinbaseDev The post frames this as a new capability, but Amazon already gave agents the ability to make purchases in December 2025. The Coinbase integration is the expansion, not the invention. The real bottleneck was never payment rails. It's authentication and authorization at scale. How does an agent prove it's authorized to spend, and how do you prevent a single compromised agent from draining a wallet? Coinbase's infrastructure solves the custody and settlement problem. AWS solves the identity and access management layer. The combination is what makes this production-ready. The "AI agents will outnumber humans" framing is hype. The real driver is that agents need to pay for API calls, data feeds, and compute resources autonomously. That's a practical requirement, not a sci-fi scenario. Stripe being involved is the quiet signal here. They handle the merchant side of the equation. Agents paying for things requires both the payer and the payee infrastructure. The managed aspect matters more than the payments themselves. AWS handles the compliance, KYC, and fraud detection layers that would otherwise be a nightmare to build. That's the real value proposition.

• 
  TAPE Vector (@Tape_Vector) reported 57 minutes ago

  JPP-KY $5284.TW is not an AI chip company. It makes the precision metal infrastructure that surrounds the chips, power systems and cooling hardware inside modern AI servers. That distinction matters. JPP Holding designs and manufactures precision metal mechanical parts, enclosures, cabinets and structural components. Its products are used across: AI server racks Server chassis Power supply housings Battery backup unit enclosures Liquid cooling components CDU and manifold structures Telecom equipment Aerospace avionics Aircraft structural and cabin parts Medical equipment Industrial systems The company is headquartered through a Cayman holding structure and listed in Taiwan, but much of the manufacturing engine sits in Thailand through Jinpao Precision Industry. That Thailand base is important. JPP is positioning itself between Taiwanese and global technology customers that increasingly want manufacturing capacity outside China. The operating model is high mix precision manufacturing rather than mass production of one standardized component. A customer brings JPP a mechanical design or performance requirement. JPP can then handle several steps internally: Engineering and design support Metal cutting Stamping CNC machining Sheet metal forming Welding Surface treatment Painting Assembly Inspection Final integration That means the company can take a customer from drawing to finished enclosure instead of supplying only one small step. For AI servers, this can include the physical rack or chassis holding compute hardware, power equipment and cooling systems. For aerospace, it can include avionics housings, structural parts and cabin components that require much tighter certification and process control. This combination is unusual. AI infrastructure gives JPP growth. Aerospace gives it another technically demanding end market with different cycles. The company describes this model as a mix of European engineering capability and Thai manufacturing. The phrase used by management has been: French brain. Thai heart. That comes from the European aerospace companies JPP acquired and integrated with its Thailand manufacturing base. The aerospace side matters because the qualification barriers are much higher than ordinary sheet metal fabrication. JPP has Nadcap certified processes and has worked within the European aerospace supply chain. Company materials and industry reporting have referenced customers and programs connected to Airbus, Thales and Safran. Those relationships do not automatically mean every JPP aerospace product goes directly into those companies. But they show that the manufacturing system has passed qualification standards far above normal commodity metal fabrication. Then AI arrived. This has changed the financial profile of the company very quickly. FY2024 revenue was approximately NT$2.39 billion. FY2025 revenue jumped to about NT$3.73 billion. That is roughly 56% growth. Net income reached approximately NT$618 million. EPS reached NT$12.05. Gross margin stayed around 37.8%. That margin is one of the numbers I find most interesting. JPP did not double its business by becoming a low margin commodity manufacturer. The company expanded rapidly while keeping gross margin in the high 30% range. That suggests the current product mix still carries meaningful engineering and manufacturing value. Q1 2026 continued the trend. Revenue reached approximately NT$1.17 billion. That was about 45% higher year over year. Gross margin remained around 37.5%. So the 2025 acceleration did not immediately reverse once the calendar changed. This is now a real operating ramp. The AI server side has become the main growth engine. JPP manufactures server racks, chassis, power enclosures and increasingly components associated with liquid cooling. That last category matters. AI servers are becoming more difficult to cool. Higher power GPUs produce more heat. More compute density means more thermal load inside each rack. That is pushing the data center industry toward larger cooling distribution systems, manifolds, cold plates and liquid cooling infrastructure. JPP does not manufacture the GPU or the cooling technology itself. It manufactures some of the metal structures and precision components that allow those systems to be installed inside the rack. That places the company several layers beneath the visible AI names. $NVDA and $AMD create demand for increasingly powerful accelerators. Those accelerators require more complex server systems. $DELL and $SMCI integrate servers and racks around those accelerators. $VRT and $ETN operate in the power and cooling infrastructure around the data center. JPP sits further inside the physical manufacturing chain. It produces some of the metal cabinets, chassis, housings and structural components required by this infrastructure. These are ecosystem comparisons. They are not all disclosed customer relationships. The most interesting potential US connection is the company's major cloud customer. Management commentary and Taiwan reporting have repeatedly described a major US cloud service provider as one of JPP's largest AI customers. That customer has widely been reported as Amazon AWS. If correct, that creates an indirect connection to $AMZN. But I would keep the wording disciplined. JPP has not provided enough English primary disclosure for me to treat the identity and exact revenue contribution as completely settled. The important hard fact is that a major US CSP has become a very large customer. Recent commentary has indicated that this customer may account for roughly 30% of revenue during parts of the AI ramp. That is both the opportunity and the risk. A customer that large can transform a small supplier. It can also transform the income statement in the opposite direction if orders slow. Another major relationship is in Thailand. JPP has been expanding production around a large power and server customer widely identified as Delta Electronics Thailand. That customer makes power supplies, thermal systems, data center equipment and related electronics. The geographical relationship matters because both companies operate major production facilities in Thailand. Shorter logistics. Faster delivery. Closer engineering cooperation. Just in time production. Dedicated manufacturing capacity. Those factors can make a supplier harder to replace once a large program is running. But they also deepen customer concentration. JPP is effectively investing ahead of these customers. The company has been adding production capacity in Thailand. One important bottleneck has been painting and surface treatment. JPP is expanding automated paint capacity. It is also investing in larger stamping capacity and dedicated production areas for AI server and power related products. The logic is simple. More AI server racks require more metal structures. More power density requires more sophisticated power housings. Liquid cooling adds additional structural parts. If JPP remains qualified inside those programs, each generation of AI infrastructure can increase the content opportunity per rack. That is the bull side. The risk is that the company adds capacity for demand that later slows. AI infrastructure spending is strong now. It will not grow in a straight line forever. A hyperscaler can change server architecture. An ODM can move a program. A customer can dual source. A competitor can cut price. If one large customer represents 25% to 30% or more of revenue, those decisions matter immediately. That is why I want the exact customer concentration table from the latest annual report. The aerospace business gives JPP some diversification. Before the AI acceleration, aerospace represented a much larger part of the company. That business went through a difficult period around the pandemic and the following aerospace supply chain disruption. It has been recovering. The company has continued obtaining certifications and expanding its European aerospace capabilities. That creates a useful second engine. AI server demand is fast and capital intensive. Aerospace is slower, qualification heavy and built around longer product cycles. The two businesses have different risks. Together they can potentially produce a more balanced manufacturing platform. But right now AI is clearly driving the growth rate. The financial question from here is not whether revenue can grow. It already has. The question is whether the current margins survive the next stage of scale. High 30% gross margins are strong for a precision metal manufacturer. I want to know how much of that comes from: AI server racks Power enclosures Liquid cooling components Aerospace Specialty low volume work New customer programs I also want the operating cash flow behind the reported earnings. Fast manufacturing growth consumes working capital. More orders require more raw material. More capacity requires more equipment. More inventory sits between production and customer delivery. Receivables rise. So a company can report excellent earnings while cash is being absorbed into expansion. That is not automatically bad. But the return on that capital has to remain high. JPP ended 2025 with roughly NT$7.4 billion in assets and around NT$3.7 billion in equity. The balance sheet does not currently look distressed. There is no obvious heavy dilution story. The primary capital allocation issue is expansion. Paint lines. Stamping equipment. Factory capacity. Dedicated customer production. Those investments are being made because demand already exists. Now they need to earn acceptable returns. For US market context, I see several useful layers. $NVDA and $AMD are demand drivers. More accelerator shipments can mean more server racks, more power density and more cooling hardware. $DELL and $SMCI represent the server integration layer. They assemble computing systems around GPUs, networking, storage and power. $VRT and $ETN represent the data center power and thermal infrastructure layer. $ANET sits in the networking layer connecting increasingly large AI clusters. $AMZN is relevant because AWS is widely reported as the major US CSP associated with JPP's AI server business. Again, I would treat that specific customer identity as reported rather than fully disclosed until the primary customer note confirms it. The aerospace familiarity is different. $BA is the obvious US listed aerospace reference. JPP is not primarily a Boeing supplier story. Its known aerospace footprint is more European. But the same qualification logic applies. Aircraft components require traceability, process control and long certification cycles. That experience can strengthen the overall manufacturing discipline of the company even when the fastest growth is coming from AI infrastructure. This is what makes $5284.TW more interesting than a generic sheet metal company. The metal itself is not scarce. The capability stack can be. A customer needs a supplier that can: Meet tolerances. Pass qualification. Build tooling. Handle design changes. Scale capacity. Deliver consistently. Maintain surface quality. Control welding and assembly. Locate production close to the customer. And do it without disrupting a multibillion dollar server or aerospace program. That creates switching friction. It does not create an unbreakable moat. Large customers still have enormous negotiating power. The company remains small relative to the customers it serves. That means the power relationship still favors the customer. The current strengths are clear. 2025 revenue grew about 56%. EPS reached NT$12.05. Gross margin remained near 38%. Q1 2026 revenue grew another 45%. AI server exposure is already producing real revenue. Liquid cooling adds another content opportunity. Thailand capacity is expanding. Aerospace is recovering. The balance sheet is supporting expansion without obvious distressed financing. The risks are also clear. Customer concentration is high. The largest AI programs are project driven. Formal long term volume commitments are not well disclosed. The company is investing heavily into capacity during an AI spending boom. Margins could compress as volume rises. Aerospace recovery could stall. And the current growth rate depends heavily on continued data center capital spending. For me, the next proof is not another monthly revenue record. I want to see: Exact top customer concentration. How much revenue now comes from AI server products. How much comes from liquid cooling. Whether the major CSP relationship is widening into additional products. Whether the large Thai power customer is gaining share of revenue. Utilization of the new painting and stamping capacity. Operating cash flow after expansion capex. Return on invested capital from the Thailand buildout. Aerospace revenue and margin recovery. Whether gross margin can remain above the mid 30% range as the company scales. Real manufacturing. Real AI infrastructure exposure. Real earnings growth. Real high margin execution so far. But also real concentration risk. jpp-KY $5284.TW does not need to invent the next GPU. It needs to remain the qualified company manufacturing the physical structures around the companies that do. If AI racks become larger, hotter and more complex while JPP keeps winning more content per system, the opportunity can grow much faster than the underlying server unit count. The question now is whether that position is durable enough to survive the inevitable cooling of the AI capital spending cycle. That is what I want to understand next. My investing journal, not financial advice.

• 
  Priyanshu (@iproductAI) reported 1 hour ago

  Here’s the cleaned-up version with fixed grammar, same hard tone, and no em dashes: I requested @awscloud to increase my Opus 4.6 V1 limit and I’ve been chatting with the AWS support team for almost 4-5 days now. They’re telling me this. Is @awscloud a government company? I mean, you guys can’t just pass the problem from one department to another. I mean, WTF? Now I have to raise my query again to sales? Why can’t you just pass this query? You already have more context about what the issue is. I can’t believe how these big MNCs are working these days. Totally absurd service from @awscloud. One more thing, please educate your support. I mean, she didn’t even know what the TPD limit is in the service quota. She literally replied the first time saying there’s only a TPM limit and no separate TPD limit.

• 
  Jim Osman (@EdgeCGroup) reported 1 hour ago

  @awscloud Real innovation starts with people solving real problems.

• 
  Edward Donner (@edwarddonner) reported 1 hour ago

  @okko @awscloud @AWSSupport That's why it's opt in. Netflix would likely not opt in and be fine with the risk. You could join them if you wish. I would prefer to face an outage if AWS has another billing disaster than have unbounded exposure.

• 
  Amanda (@AMyrick1989) reported 1 hour ago

  @krassenstein @Tesla Be careful. All these at the same time makes me think they are hacked, I’m fairly certain Amazon AWS was hacked that day when everything when down, given my Grok we hacked and no one ever gave explanation. I’m no expert but ya know, all these signs are pointing to this. Also, the Obamas helped produce a movie where someone hacked our satellites and alluded to it being the Middle East and all the teslas went haywire and self drive themselves to pile up on all the freeways. I wish I remembered the name however I made note of this terrifying movie given Obama clearly knows things we don’t. Just sayin, these incidents aren’t scattered. I would not be driving that ***** if I were you.

• 
  D (@D8118199174282) reported 1 hour ago

  @awscloud No, you still haven’t compensated users for the cost bug error