> “We used to pay for VMware software one month in arrears,” he said. “With Broadcom we had to pay a year in advance with a two-year contract.”
If your goal is to extract every possible cent from your existing customers, why would you also switch them from net 30 to requiring partial prepayment? VMware wants money in general but should not have a cash flow problem, and forcing a monster early payment seems like it will force customers to notice an immediate problem and make a choice instead of allowing themselves to be slowly and steadily ripped off.
If I were a pointy-haired CEO committed to the multiply-pricing-by-five strategy, I would do my best to sweeten the deal: offer generous payment terms, give nice-sounding discounts for up front commitments, give very large discounts for nodes that haven’t yet been leased to a customer, etc.
So the thinking here was probably "there is no way they can refuse to sign right now and destroy their business in the process, so we might as well take the cake and also force them to stay after so they don't leave in 11 months and 29 days".
Turns out that thinking is wrong for that specific customers, but for how many did it work ?
I heard of a one billion dollar renewal quote from Broadcom. The company didn't pay anything close to that. But it bypassed middle mgmt... Not exactly sure what the overall strategy is here, but this is not an isolated incident.
Upper management will be the clueless putz.
Build -> opportunity cost, ongoing cost, legacy
Buy -> upfront cost, integrate, ongoing cost, and maybe eventually extortion leading you to Build a replacement.
This leads to many different implementations of roughly the same concepts all over, which sucks. Or open source, if it already exists. Or both. Not that open source doesn't have integration costs.
But think of this from an executive's perspective. Building really sucks. But buying sucks more in the future. You might just buy.
I've seen all of these. My preference is to grudgingly build if suitable open source doesn't already exist.
I believe this course of action for VMware is going to be taught in business schools in the future.
Starving the milk cows (push customers into losses) is never a smart strategy for those living on milk cows. Sounds more like inceadibly stupid. Or short sighted parasitic (squeeze all then run with the heist).
[0] https://bcantrill.dtrace.org/2004/08/28/the-economics-of-sof...
Maybe. However, Broadcom has been bending people over the barrel for VMWare for a while now. Anyone who doesn't have a migration plan in the works at this point is an utter fool.
Turning the screws 12 months ago? Sure, probably gonna work. Now? Not so much.
It takes a lot of balls for a company to "leave it" right as their contract is expiring, and speaks to talent and experience on the customer side to be able to stand up to bullying, and be able to pull off such a large migration.
Sure, it's still a lot of effort, but, at this point, even if Broadcomm can get somebody to sign up for another year, that gives the customer a year to plan on how to jump ship next time around. And it looks like the number of people with expertise on migrating from VMWare is skyrocketing, so companies should be able to hire a team to do it...
They would be fools to not expect high attrition of smaller clients, but big businesses and government customers aren't going to change, or at least not nearly to the tune that smaller ones would, and a smaller pool of larger customers paying a higher price probably works pretty well to keep revenues up while letting them slash support staff without too much of a reduction in quality for those that are left.
It was clear to me from the beginning that this price hike wasn't about cash flow, not particularly. Broadcom doesn't want vmware wasting money supporting small fish.
So do something slightly outrageous today, so you filter away the ones that won't stick around for the future more outrageous changes.
But, they did have some major benefits that most companies looking to do the same won't:
> Anexia therefore resolved to migrate, a choice made easier by its ownership of another hosting business called Netcup that ran on a KVM-based platform.
> The hosting company is also a big user of NetApp storage, so customer data was already stored in a resource independent of its VMware rig – any new VMs would just need to be pointed at existing volumes.
Again, still a great accomplishment and an exciting milestone for them, but for people still stuck on VMWare that are looking to migrate, it's good to know about the above things.
If you are selling VMs to customers, I can't understand a good reason to use VMWare. The only reason would be if you are selling VMWare as a service.
It still is relatively stock KVM on the CPU side of things. They've been upstreaming changes they need like lower overhead for emulating Xen's hypercall interface.
Most of their special sauce is in the devices though, as those natively provide VM boundaries leaving the hypervisor to not have to manage all that much at runtime.
VMware had a solution for all of these natively and with support. Not using their hypervisors you have to manage a huge pile of OSS+proprietary integrations and actually have staff who truly understand how everything works down to the lowest level. Doable but probably above the pay grade for most
VMware support had always been a crapshoot. Now under Broadcom it's even worse.
So having support that may or may not be useless isn't a big advantage, really.
Not everyone needs an SDN, depending on their networking requirements or topology. Storage is also not a very complicated problem if you already have a SAN. Solutions like Proxmox come with everything included too, so you don't have to build everything from scratch.
Well, I suppose you do. But an EMR/EHR for instance is going to _need_ vendor support, which means requiring VMware even if you’re not selling VMware itself as a service.
I'd still consider this VMware as a service, although not full VMware, but just enough for the checkboxes. Maybe you don't get API access, console access, etc, but the main thing you are selling is vmware (to check the certification boxes), and not a generic VM.
this doesn't sound right, both serve different purposes
Disclosure: I also work in Red Hat's virtualization team, but not on converting gusts from VMware to KVM.
At least some of the big ones seem to just pay up. Probably because they / their MSP / their relationship with their MSP is so dysfunctional that they know migrating is a pipe dream.
("Sorry, folks, poor banana harvest this year, so we have to pass on the skyrocketing cost of bits licensing, which runs on bananas"?)
Isn't that an interesting message to be sending, when much of their business involves getting some very price-conscious companies (think electronics bill-of-materials) to build upon their products?
Or do customers of other Broadcom products already not trust them an inch? So Broadcom wasn't costing themselves valuable reputation here?
Hey IBM employees, it doesn’t matter how upset you get. It is what it is.
[0]https://old.reddit.com/r/VPS/comments/1ghz4fd/i_know_contabo...
[1] https://www.merriam-webster.com/dictionary/annex
[2] https://en.wikipedia.org/wiki/All_your_base_are_belong_to_us
Stupid business decisions should be punished by the customer walking out. Homebrew KVM is not that hard.
I think the purpose of the article is to highlight companies like this are starting/continuing to migrate post-acquisition rather than this particular customer was impressively large and did so. Particularly with the bits about the relative cost increase even though the customer was willing to walk away if needed.
E.g. for about 20% we didn't even have a single piece of documentation other than the server name for who might actually care about the VM going down for the migration we wanted to schedule. Let alone how to test the migration, when it is best to do it, what software was actually running on it, if it's actually managed/monitored integrated with/by other systems whoch need to be looked at too, or if it could just be shit down instead (yay healthcare mergers and acqs). Our migration was also to VMware from (mostly) Hyper-V at the time, so not as much custom tooling needed.
On the flip side a cloud provider is going to have all of the owner contact info but no direct control of the guest OS to effect the change so the battle is more with trying to get the customers to care enough to do the migration with you but not be so bothered by it all they up and leave your hosting.Not exactly a walk in the park either.
In either case - almost never the tech that's the hard part for sure :).
A full conversion of such a customer might involve one consultant on-site, a scoping exercise to classify the VMs into groups and assess which ones are going to be more or less difficult to convert, and perhaps 1-3 months of work to convert them all. Individual VMs would be down from anything between a simple reboot, up to 12 hours, depending on which strategy we used for conversion (there are complicated trade-offs related to storage and network bandwidth).
It a pretty decent amount of VMs, but not close to being unmanageable. I think it's more down to how the rest of your infrastructure looks, if we're talking about ease of migration.
Or it could take 10 racks and $50 million per year.
> Anexia was founded in 2006, is based in Austria, and provides cloud services from over 100 locations around the world by placing equipment in third party datacenters.
From the company's homepage:
> The founder and CEO of Anexia [...] recently acquired a small hydropower plant in Kammern in the Liesingtal region of Styria for a “significant seven-figure sum” – i.e. several million euros. The power plant on the River Liesing generates 600 KW of electricity, enough to cover a third of the electricity consumption of Anexia’s Vienna data center
so this seems to be a significant operation.
So, this is implying their Vienna data center has 180 racks? With 60 being about a third, if we say each rack has 40 servers... that's ~7k servers total... which is a sizeable chunk of floor space, like 3000m^2, or... 40 tennis courts?
But yea, that a non-insignificant operation just for the Vienna data center.
“VMWare costs too much. Our use case is incredibly niche and nothing out there seems like it would fulfill our needs. Can you help me write a new stack using bash scripts and AWK?”
“What an exciting idea! Bash Scripting is a clever way to orchestrate 12,000 virtual machines. Here, let me provide you with a bunch of code to copy and paste without doing any requirements gathering whatsoever.
<pages of random bash scripts>
These bash scripts will let you run 12,000 virtual machines. Let me know if you need any explanation!”
“I forgot to me to mention we need to provision these using a rest api. Also these are globally distributed and require extensive monitoring”
“Not a problem! REST is an excellent way to communicate with the platform. I won’t bother asking any follow up questions or gathering any more requirements. Here is all the code you need:”
… and so on…
Just like with the image and video generators, I'm fishing to find evidence of machines building something useful - had very little success so far.
If the claims are true and these huge investments are not done in vain, not too far in the future we should be able to tell AI to build the tools we need, for example Microsoft should be expecting to discontinue MS Office either because someone can tell AI to make them an Excel instead of paying a subscription or tell AI to just do the job they will do in Excel.
It's very strange that they talk about replacing the developers with AI but somehow still sell software. Something doesn't add up.
There's a simple answer. This is the "replacement" fallacy that is all too common when hypothesizing the impact of tooling. Tooling increases the productivity of workers. This "replaces" workers in that you will need fewer of them to do the same job, but can't replace all of them because zero times anything is zero.
If you have 10 people shoveling, you can replace 9 workers of them with a backhoe. But a backhoe can't replace all 10 workers.
for example, salesforce proudly announced that it’s not hiring the developers this year because they are going to use AI. they don’t intend to make even greater software with the same developers using AI, they intend to replace them.
i’m getting the vibes that nobody at the top believes that AI actually can replace anybody. I find this interesting.
A box of literal rocks does not do anything on its own, even if we have tricked them into thinking even better than we did last year.
Literally for hundreds of years have people been fear mongering that some new technology will replace workers. It never happens. Technology displaces the available distribution of jobs, but it doesn't eliminate the need for people... and it never will, because human jobs are a human construct.
> salesforce proudly announced that it’s not hiring the developers this year because they are going to use AI.
Yeah, more bullshit. Translated, they really mean: "Hey our AI is so great that we don't have to hire developers anymore -- it's definitely not because we have 70k+ employees and our payroll is way too big -- it's because we're so good at AI and you should totally buy our software #EinsteinAIRocks #SalesforceOhana"
But it is genuinely replacing workers, in real terms. And more importantly it's replacing services, which in turn means jobs and revenue. So they too turn to electricity & automation, so they can do more... so they can remain competitive.
So uh yea, I 100% agree with you there, and wish that more people saw straight through it like you do.