AWS EC2 General Price Cut | Hacker News Reader

AWS EC2 General Price Cut

aws.amazon.com

137 pointsmakeshifthoop110 comments

Buying your own Physical hardware is so cheap now its crazy. A commodity Ryzen 16 vCPU box is about the same price as 2 months rent on EC2.

I get maintenance, electricity and bandwidth isn't free but I honestly thought cloud server prices would be much lower by now. No wonder AWS is making huge profits.

For me, not having to worry about the "where" and "how" related to physical placement and connection of that Ryzen box is worth the premium and then some.

Then use a dedicated provider. You won't save as much, but in raw costs, you'll be about 2x ahead and in cost/performance you'll be 4x-8x (much more if you're a heavy bandwidth user, and much much more if proper hardware lets you scale vertically rather than prematurely having to scale horizontally).

AWS is so much more than just hardware

It is called service for a reason.

With AWS you also pay for much more than just hardware.

Even when compared with other cloud providers, AWS is super expensive.

The only interesting service that AWS provides is the ability to host your services globally through multiple availability zones, but unless you need to provide a global service with low latency everywhere in the globe then that's just not worth the time and money.

> Even when compared with other cloud providers, AWS is super expensive.

Well this is not true. I don't think Azure is much cheaper in that regard. In reality, the big clouds are watching each other very closely. Do you have data other than anecdote evidence?

Also if you are big enough, you can negotiate better deal with AWS separately, and I believe that is the norm.

Disclaimer: I worked at AWS before.

> Do you have data other than anecdote evidence?

Why, yes. Take Hetzner for example. With their Hetzner cloud offering you get, say, 1vCPU with 2GB of RAM and 2TB of traffic for 3€/month. For an instance with 2vCPUs yo pay about 4.15€.

For an equivalent EC2 instance (t3.small) you pay about 5x the price, and you still need to pay additional costs such as egress charges.

Heck, with Hetzner you can spend less than 50€/month and get a dedicated box with 64GB of RAM and 4 real cores with free unlimited traffic over a dedicated 1Gb connection.

It isn't even up for debate: AWS price-gouges their customers. Unless you have a very particular need to deploy an application globally, it's very hard to defend paying AWS's cost.

The Hetzer-equivalent service is Amazon Lightsail, which starts at $3.50 a month ($5 for 2TB transfer) and has access to all AWS services.

> The Hetzer-equivalent service is Amazon Lightsail,

No it isn't. Hetzner provides full blown linux instances with 2GB of RAM, while the Ligthsail instance type you're trying to compare has a barely workable 512MB of RAM. That's the equivalent of a meager Orange Pi Zero.

The Lightsail instance type that is equivalent to Hetzner's 4€/month instance is sold for 20€/month.

Check the prices for example on Hezner or some other company which provides dedicated boxes. They are totally different from big cloud. Like 6 cores and 256GB memory for 150€/month (one older server I have there).

Of course these don’t scale dynamically. And you don’t just create copied of the servers. And they don’t have 100 security certifications like Azure/AWS. And no data centers in every jurisdiction.

> Like 6 cores and 256GB memory for 150€/month (one older server I have there).

As I've mentioned in another post, Hetzner also sells for less than 50€/month dedicated hosts with 64GB of RAM running an Intel i7 6700 professor (4 cores, 8threads).

With that you also get unlimited network traffic over a dedicated 1Gb connection.

Vultr has servers with pci certification.

Could someone a little more familiar with buying their own hardware tell me how much hardware equivalent to a c5.12xlarge (48 cores, 96 GB) would cost?

Spot prices are around $0.80/hr, which puts total compute cost at $7k/yr.

On demand is $2.04/hr ($17k/yr), reserved is 1.20/hr ($10k/yr).

Not colocated, but in the US, you can find a dual EPYC 7742 (64 cores) with 512GB of ram (with ssds, hdds, and bandwidth included) for $1000/m.

This should absolutely demolish the c5.12xlarge, especially if you need any disk IO or bandwidth.

Or from a different provider, a Dual Epyc 7281 (32 cores) with 256GB (and ssds, hdds and bandwith included) for $400/m.

You can get cheaper if you're willing to host in Europe. And this is dedicated, so no additional power or storage or maintenance cost. Stuff breaks, they replace it. Also, your commitment is 1 month.

To add context to the downvotes you've acquired: I didn't downvote you, but I think I understand why others did. Your parent comment was asking about 96 GB of RAM, and you posted a reply talking about 512 RAM. I suspect others thought you were talking about 512 MB (ie. 0.5 GB of RAM, which is common in very cheap hosting) rather than the 512 GB of RAM you actually commented on. When I first read your comment, I admit I first saw "512 MB", but you quite clearly wrote "512 GB".

tldr; The number "512" when talked about in the context of RAM is assumed to be measured in MB, but 512 GB is also a less-commonly-encountered amount of RAM. :)

Current gen hardware is about 10-15k. If you buy used or legacy hardware (like what c5 runs on) you can get much more bang for buck.

Here's an eBay listing for 2 machines with similar CPU config for 400+shipping: https://www.ebay.com/itm/INTEL-COMPUTER-48-CORE-DUAL-XEON-24...

Buyer be wear but there are deals to be had. That said purchasing the hardware is step one. Setting up PXE, network storage, cluster mgmt, recycling and upgrade, etc is a full time job passed a specific scale.

Not exactly what you asked for. But, we bought an AMD server to test at our HPC cluster. AMD EPYC 7302P (16cores, 32 threads), 512 GiB of 3200 MHZ RDIMM, 10GBE, 3.2TiB of 12Gbps SAS SSD. It came out to be approx 48000 DKK (excluding VAT). That is approx. $7000. Fewer cores machine with faster clocks and we mostly use it for its memory channels(8).

Our typical buy - 2 socket intel cascade lake Xeon Gold 6256(12 Cores, 24 threads) with 768 GiB of RDIMM costs us approx. 70000 DKK or $10200.

The biggest cost here is memory. Just to give you an idea.

I was just thinking the other day when HN were discussion Cloud vs Metal. How little has the DRAM prices changed in the past 10 years. While CPU core went from 10 to 64 Core, ( Not exactly Moores's law at least it is 6.4x the difference ). DRAM median prices has remain mostly flat until recently, we finally got pass $10 /GB ECC RAM and are now close to $5/GB. Which is still very expensive.

SLC NAND SSD could be had for $0.4/GB, I failed to see why DRAM should be priced at 10x higher.

mhm, interesting question... i am not dabbing in the server space just in consumer hardware...

1x ryzen 3990x with 64 cores(128 vcores) would be 4k € 1x a mainboard ~0,5k € 1x 128 GB RAM ~ 0,7k € Storage, whatever you need 0.5k case + power supply 0.3k

so 6k all around for a much higher perfomance pc...

a ryzen 24cores(48vcores) ~ 1k 1x a mainboard ~0,3k 1x 96GB Ram ~0,5k Storage, whatever you need 0.5k case + power supply 0.3k

so 2.6 k should be around the same in raw consumer hardware

Not the same as buying hardware, but Hetzner offers dedicated AMD Epyc 7502p instances (32 cores, 128GB RAM, free egress) for less than 150€/month.

So you can just get somewhere between 5 to 7 of those instances for the price you're paying for your single EC2 instance.

Yeah for big iron EC2 makes more sense. Still you can find similar eg this refurbished for 6k https://outlet.us.dell.com/ARBOnlineSales/Online/SecondaryIn...

It’s buy vs build but with people being routinely misled about how much it costs to build and the fact that it doesn’t have to be all or nothing.

Some companies are making billions of dollars, that’s why we see so much content about the cloud and very few about building your own cloud with dedicated providers, which by the way handle electricity, network and basic hardware maintenance for pennies compared to cloud guys. They even have some basic managed services like backups, S3-compatible storage etc. these days.

PR speak aside, it looks like they are trying to get people to switch to certain regions, perhaps which are not popular with lot of idle machines? The popular locations like Oregon and Virginia don't have much discounts to offer.

Makes sense, trying to mixmax their load by adjusting pricing for people who care less about region and more about cost.

for the US regions, california is already ~25% more expensive, so this just reduces the difference

Are they still extorting customers by price gouging bandwidth? Maybe one day I’ll wake up to news they’ve decided to charge for capacity instead of transfer.

> Are they still extorting customers...

That is a contradiction in terms.

According to the OED:

Extort means to "obtain (something) by force, threats, or other unfair means."

Customer means "a person or organization that buys goods or services from a store or business"

Thus, extortion is a coercive relationship--we call the target of extortion a victim, not a customer.

The relationship of a vendor and customer is voluntary.

It's nice to see that the increasing amount of hyperbole here isn't impacting everyone.

It's not just hyperbole, it's grossly unjust.

Amazon quite rightly charge a higher price for some portions of AWS--transparently, up front, with tools to assess usage and estimate fees beforehand--because they created a product that many organizations believe provides extraordinary value.

And for the sin of earning a profit through voluntary trade, a vocal minority of HN damns them every single time AWS is mentioned. Every time we let this evil behavior go unchallenged, all productive individuals are diminished.

Wait, are you saying it’s grossly unjust to criticize the pricing of Amazon’s cloud product?

Just to be clear: Amazon is not paying a variable cost for how many bits are transferred over their wires. And the standard, prior to cloud, and still in colocation, was to charge for capacity (i.e. $/gbps) rather than transfer (i.e $/gb). They are making massive, massive profits by this pricing arrangement. The cost of sustaining 100 gbps transfer for 30 days on Amazon is orders of magnitude higher than paying for one month of 100 gbps of IP transit.

They’ve normalized a pricing structure that is disconnected from any actual cost basis. You’re free to pay for it, but it’s quite amusing to say criticizing that is unjust.

"[...]disconnected from any actual cost basis[...]"

Not everything is priced based on the cost of it's inputs. In this case it seems to be what the market will bear.

Which in this case is exorbitant and is something that we should oppose. Pricing as much as the market will bear is not good for consumers, and in some cases is a red-mark of market failure (not in this one, as of yet).

> Pricing as much as the market will bear is not good for consumers

To be fair, that's how all products are priced in general: production cost defines the lower bound and what customers are willing to pay defines the upper bound, and the goal is to maximize what the customer pays.

The theory is that as time goes on, the prices moves from what the market will bear to become closer to the cost of production. When this does not happen, it is a failure of the market.

Terms are commonly used metaphorically to show the speaker considers a given situation to resemble that term qualitatively but not exactly. Very commonly. Like, it's practically fundamental to our communication.

You can coerce someone to be your customer.

The grandparent's argument is that AWS is extorting the customer to remain a customer by making the cost of moving data out of AWS prohibitevly high. I don't think it is actual extortion (at least legally), since the egress prices are public, and you could argue if the customer had done due diligence they would have known the price when becoming a customer in the first place.

But if they had changed the price after a customer became sufficiently invested in AWS, that would be using "unfair means", to obtain the customer's continued business.

   220 pan.alephnull.com dictd 1.12.1/rf on Linux 4.4.0-1-amd64 <auth.mime> <114924196.15062.1589085014@pan.alephnull.com>
   150 15 definitions retrieved
   151 "Extortion" gcide "The Collaborative International Dictionary of English v.0.48"
   Extortion \Ex*tor"tion\, n. [F. extorsion.]
      1. The act of extorting; the act or practice of wresting
         anything from a person by force, by threats, or by any
         undue exercise of power; undue exaction; overcharge.
         [1913 Webster]
   
      2. (Law) The offense committed by an officer who corruptly
         claims and takes, as his fee, money, or other thing of
         value, that is not due, or more than is due, or before it
         is due. --Abbott.
         [1913 Webster]
   
      3. That which is extorted or exacted by force.
   
      Syn: Oppression; rapacity; exaction; overcharge.
           [1913 Webster]
   .
   151 "extortion" wn "WordNet (r) 3.0 (2006)"
   extortion
       n 1: an exorbitant charge
       2: unjust exaction (as by the misuse of authority); "the
          extortion by dishonest officials of fees for performing their
          sworn duty"
       3: the felonious act of extorting money (as by threats of
          violence)
   .

"an exhorbitant charge"

Actually, I do genuinely wonder if some day this will be seen as anticompetitive behavior (a form of lock-in via higher than reasonable egress fees).

Anticompetitive behavior when they clearly state the price? Of all the things that lock you in to any provider once you are at any scale, data isn’t the biggest one.

If you have a massive amount of data, you can always get AWS to transfer it to you via a Snowball for $30/TB + a $300 fee.

IANAL, but I don't believe stating a price has any impact on whether it is anticompetitive or not. If the behavior is anticompetitive, it doesn't matter if other people do it too and state their price as well.

Giving you an option to only bulk egress is not the same either, since the most common scenario for egress would be operational day-to-day inter-network scenarios.

If there is an unreasonably high fee for exiting a provider's cloud, to any other network location, on an operational basis, it is actively discouraging exiting the provider's services creating a kind of price barrier that could be cost prohibitive for operational inter-networking.

Operational inter-networking, to me, encourages competition not discourages it.

For example, perhaps I want to use Microsoft Azure's Data Factory to move data from AWS S3 to Redshift. It is not possible as far as I know without egress from AWS to Microsoft's network, and then back again. But surely these clouds are connected with peering relationships making the cost of ingress/egress negligible, so why is this not allowed? That would allow users of the cloud to freely mix and match services from among many clouds.

The world I see today does not seem to encourage such competition.

The Bandwidth Alliance was meant to solve this problem. Azure and GCP have joined, AWS has not. https://www.cloudflare.com/bandwidth-alliance/

Wow, I hadn't seen this. This is quite cool. So clearly somebody else at Cloudflare saw the same concern I do and came up with a creative solution. Interesting that AWS isn't playing though.

That’s a classic move for second tier players to come up with a standard to take on the incumbent.

Apple did the same thing with Webkit. Google came onboard and for awhile, everyone was releasing Webkit compatible browsers.

One factor could be while of course they are peered it might be paid peering or if they allowed the traffic unmetered it would upset other free peering arrangements due to subsequent asymmetrical transfer rates.

> you can always get AWS to transfer it to you via a Snowball for $30/TB + a $300 fee.

1. That fee is in addition to S3 retrieval costs.

> Q: What does it cost to export my data?

> In addition to the Export job fees detailed on our pricing page, you will also be charged all fees incurred to retrieve your data from Amazon S3.

https://aws.amazon.com/snowball/faqs/

2. $30/TB isn’t even cheap by itself.

If you’re any kind of business ,$30 a TB is cheap.

S3 request fees are $.0004 per 1000 GET requests. If you had 100000 objects (files) you would pay $40

the best form of competition will be simpler distributed compute, with the ability to more seamlessly process workloads across end user device, edge compute and cloud.

of course, AWS will play there too, but doesn't mean the overall ratios won't tilt towards the edge (although plenty of data "left" for the clouds since overall data will continue to grow).

I suppose, like what is happening with Kubernetes.

I thought anticompetitive behavior was the exact opposite of price gouging. You sell below cost to put your competitors out of business. Evidently AWS isn't doing that with bandwidth.

If you price lower than your competitors it is anti-competitive dumping.

If you price higher than your competitors, it is monopolistic gouging.

If you price the same as your competitors, it is price-fixing.

No matter what price someone selects, they are guilty of something.

Well no, it's only gouging if you're running at a loss. If you're not running a loss and pricing lower than your competitors that's just fine.

Egress is expensive, therefore you're less likely to use other cheaper cloud vendors to host some part of your infra and keep everything inside aws network (as internal aws bandwidth is free even halfway across the world). Also if you have huge amount of data, moving them out of aws into another cloud would net you a hefty bandwidth bills so you're less likely to consider moving.

Not true (unfortunately). Transfer within an AZ is free, but go outside that and the pricing is a minefield:

https://twitter.com/QuinnyPig/status/1172239124251709449?s=1...

Ah you're right, I somehow remember it as free.

AWS bandwidth fees make a multi-cloud infra a less likely design or migration choice by their customers, leading to lock-in.

> Service providers may attempt to “lock in” customers to prevent them from switching to alternative products, technologies, or suppliers. Customer lock-in involves raising customers’ switching costs to the point that the cost of switching outweighs the potential benefits from switching.

http://www.ictregulationtoolkit.org/toolkit/2.2.3.6

And traffic bandwidth?

I know it is not Christmas yet but it would be great if those prices trend to marginal cost for the sake of innovation.

Dont be silly.

Have traffic prices even moved in years? It’s such a blatant ripoff.

Good for backups though.

Write once, read never!

I'm waiting for the day they announce bandwidth price cuts

this. the only way I can make profit with my web service is by using a third party low budget CDN as a caching layer. Even with that, S3 data transfer remains my biggest cost. if I could spare the time, I'd make a custom caching layer so that S3 is never hit at all except for upload, and is used only for disaster recovery.

Sounds like you just had a million dollar idea. Why not build it out and offer it to the many many other people in your position?

You don't need to be a plumber to spot a leaky pipe

unfortunately my current and income-giving project takes priority. but if someone else makes it, I'll use it for sure.

Backblaze B2 is what you’re looking for.

those prices are certainly nicer!

And they are in the bandwidth alliance with cloudflare

I haven't done the math, but depending on your bandwidth requirements - you could try spinning up a couple vms at DigitalOcean/Linode/somewhere with bundled bandwidth, and then just run nginx as a cache with s3 as the backend.

AWS Lightsail is a possible workaround for bandwidth heavy simple workloads. The smallest $3.50/mo instance (t2.nano behind the scenes), comes with 1TB outbound bandwidth - equivalent of about $90 for the same in EC2 excluding the instance cost, and bundled bandwidth increases for larger types.

I wouldn’t risk AWS closing the account. It should work, but don’t know how would AWS view it.

"You may not use Amazon Lightsail in a manner intended to avoid incurring data fees from other Services (e.g., proxying network traffic from Services to the public Internet or other destinations or excessive data processing through load balancing Services as described in the Documentation), and if you do, we may throttle or suspend your data services or suspend your account."

That reads to me like an admission that they know their normal bandwidth charges are ridiculous.

"The price reduction you receive, depends on the region you choose, whether you take out a 1 or 3 year term and finally the instance family you commit to in your agreement. Price reductions vary from between 1% to a massive 18% on what you were previously paying"

18% reductions for 1 year terms in Europe.

8% in mnt US region for 1 year term. No savings elsewhere.

Some prices reductions elsewhere in the world.

This is far from a general price cut. It's a price cut only of you are willing to commit to long term usage.

I'm holding out for a better price on GPU instances (for parallel processing).

My 'back of the napkin' calculations indicate it would take very little time for it to pay for itself by purchasing your own NVIDIA GPU(s) and running them from a colo.

Especially if your requirements are for the instance(s) to run 24x7.

Why would they ever be not the case ? If you ignore sysadmin costs AWS is always more expensive (even when reserved for three years) than any machine you maintain yourself

I do my own sysadmin work when needed, but I find plenty of places where I'm gladly willing to make the relatively small concession (pay the cloud markup). In the case of GPU instances I find they're not even trying to be competitive with market costs.

Isn't the whole point of cloud that you don't own the machines? The calculus probably isn't as good when you factor in support, maintenance, and power costs right?

Power is accounted for in my calculations. Technically today's gpu compute does wonders for power consumption. And likely to only improve from here.

In the many years I've used cloud I've never requested support, so in my experience I don't need to calculate that in.

And even when you consider machine costs, as long as you can get at least 1 yr (not an unreasonable assumption) out of your equipment you're more than making up for what it would've cost in the cloud.

Can you actually buy GPUs now? An advantage of Amazon in the past was that they were actually available there, where they weren’t available to buy in practice.

The crypto scams seem to have died off, so there are probably plenty of used ones on the market.

The demand right now has also fallen dramatically even from regular customers.

https://www.tomshardware.com/news/coronavirus-motherboard-gr...

oh, yeah. Thanks for reminding me. Disclaimer: My needs are not for crypto currency :)

Not sure what you mean. You can buy GPUs anywhere.

You can even buy Tesla V100s that power AWS accelerated instances.

> Not sure what you mean.

Over the last decade there was often a major problem in GPU supply. Huge users like Amazon were presumably buying in bulk direct from manufacturers as custom runs so weren’t trying to buy them on the open market like everyone else.

Even other clouds had major problems getting them at some points.

There were lots of GPUs being bought up for crypto mining reasons, but that doesn't seem like it has been as big of a deal in the consumer market recently (at least locally.)

If anything, it generated a lot of cheap secondhand cards for sale in some places.

There was never a real issue getting 4 or 5 commodity-grade GPUs. At worst, you had to pay a 20-30% premium.

No change in pricing for GP2 in six years.

Hmm, given the uneven reductions between regions, I wonder if this is more of a move to balance (and increase) demand so that it more closely matches their supply. So then I guess the question is: why do they have excess supply?

Competing cloud service providers gouging away at what used to be an AWS monopoly that unfortunately for Amazon, no longer exists.

Alternatively, some regions are cheaper to add capacity to than others (either cap-ex and/or op-ex), so extra capacity gets steered there.

I'm wondering if this is to support small businesses/startups in the light of Covid19. I know they amazon and google are already giving covid19 discounts to companies but that won't cut it. I personally know so many startups who are looking at their huge server bills and won't be able to survive for long if that pricing continues. A failed startup is just one less customer for cloud companies. So could this new pricing be to to keep existing small/medium customers ? Just a thought. What do you guys are experiencing around ?

Can someone recommend dedicated server providers with US, EU and Asia locations? Hetzner looks great but they only have EU location.

OVH?

I often wonder why Hetzner dont expand internationally.

AWS and EC2 is probably a pretty good deal for high CPU, hi IO within the cloud, but very low IO egress.

Monte Carlo simulations? PDE computations?

Wouldn't a lot of that stuff be better done on GPU instances, even at a higher per hour cost?

EC2 also has GPU instances, right? The p2/p3/g3/g4 series.

This raises a good point. Would anyone be willing to give some examples of use cases where egress to cloudfront or ec2 isn't viable? Most useful computations I can think of involve consuming large amounts of data and simplifying it to some easily digestible result. Media streaming is the first and only example that comes to mind, but I'm sure there must be others.

They're cutting only Intel instance-type pricing, a signal that AMD and ARM instances are relatively more valuable?

If I just purchased a few RIs 2 weeks ago, did I miss out on the savings, or are they somehow retroactive?

Reach out to your account manager.

The price reduction for M5 in US East is just 2%. Wish it was something that was actually substantial.

Noticed that too. We just upgraded all our servers from M3 to M5 instances, but all in the us-east-1 region, so looks like we won't save a lot by purchasing more reserved instances.

Noticed a huge 20% reduction in some other areas though...

The us-east-1 region is widely considered as the "default" region, and so it's where customers put resources when they don't have any particular reason to put it somewhere else. For that same reason, it's also where less-sophisticated customers are likely to run unreplicated workloads. It's also the oldest, and biggest, and home to the most different instance types and rack architectures. It makes life hard when your largest, oldest, weirdest, and most important region are all the same region, so AWS has been trying to nudge customers off of old instance types and out of us-east-1 with pricing changes for a long time.

How does this compare to a general macroeconomic GDP forecast over the next 3 years?

I don’t know, but if AWS prices reflected economic reality they should reflect the GDP hit and the shift from real world to digital services. It’s unclear which one wins.

Still expensive

Will Google Cloud follow up?

They basically did a while back with the E2 series of instances.

https://cloud.google.com/blog/products/compute/google-comput...

Google already did. They offer significant costs savings over AWS. Google also allows unlimited egress to other Google services.

Reply on news.ycombinator.com