A decade ago, ‘cloud first’ was the safe answer for every media and technology team. Today the more useful question is narrower: which workloads belong in the public cloud, which don't, and can we move between the two without a rewrite? That is what a cloud exit strategy is for.
An exit strategy isn't a decision to leave. It's the ability to leave on your own timeline and for your own reasons, whether that's cost, sovereignty, performance or negotiating leverage. Broadcasters, streaming platforms and AI teams feel this most acutely, because their workloads share three traits the hyperscaler pricing model punishes: steady utilisation, enormous data gravity and heavy egress.
Why this is on the agenda now
Several forces have converged in the past two years.
- Regulation. The EU Data Act's cloud switching rules have applied since 12 September 2025, and from 12 January 2027 providers can no longer charge customers for switching, including the data egress the switch requires.1 In the UK, the CMA concluded in July 2025 that competition in cloud is not working well, citing egress fees and licensing among the barriers.2 Portability is moving from a nice-to-have to an expectation.
- Providers have blinked. Google Cloud, AWS and Microsoft Azure now waive data-transfer fees for customers moving their data out, subject to approval.4, 5, 6 That removes one of the biggest one-off obstacles, though not the ongoing egress fees you pay while you stay. The Data Act ban covers switching, not everyday traffic.
- The GPU economy. AI training and inference have made compute the single largest line on many cloud bills. Long-running GPU clusters behave nothing like the bursty web apps the cloud was priced for.
- Public proof points. 37signals, the company behind Basecamp, reported saving almost $2 million in its first full year after moving compute off AWS, and projected total savings of well over $10 million over five years once its 10 PB of S3 storage followed.7
Key dates
The timeline that has turned cloud portability from a theory into a planning deadline.
-
Google Cloud waives exit egress
Free data transfer for customers leaving Google Cloud, after approval and within a 60-day window.4
-
AWS and Microsoft Azure follow
Both offer free data transfer out for customers moving to another provider or on-premises.5, 6
-
AWS cuts GPU prices by up to 45%
On-demand pricing for H100-based P5 instances falls by 44%, resetting the rent-or-own maths.8
-
CMA concludes its cloud market investigation
The UK regulator finds competition is not working well and names egress fees as a barrier to switching.2
-
EU Data Act applies
Switching obligations begin. Providers may only charge reduced, cost-based switching fees during a transition period.1
-
CMA action on Microsoft and AWS
A strategic market status investigation opens into Microsoft, and both Microsoft and AWS commit voluntarily to reduce egress fees for UK customers.3
-
Switching charges end in the EU Upcoming
Providers can no longer charge for switching, including the egress the switch requires. Existing contracts may need amending to benefit.1
Rent the spike. Own the baseline.
The principle behind most successful hybrid strategiesBroadcast and streaming: where the bill hides
Media workloads are a mix of the predictable and the spiky. A 24/7 linear channel is the most predictable workload there is. A cup final or a global premiere is the opposite. Treating them the same is where costs run away.
Egress is the real tax
Every byte that leaves a cloud region to a CDN, a partner or a viewer is metered. For a streaming service delivering millions of hours a month, egress can rival or exceed compute. Owning your origin and peering directly with CDNs and ISPs can turn a variable, per-gigabyte cost into a fixed port cost.
What 1 PB a month of egress costs at list price
$53,800 a month
About $646,000 a year, for one petabyte a month leaving a single AWS region for the internet.
| Tier (per month) | Rate per GB | Cost |
|---|---|---|
| First 10 TB | $0.09 | $900 |
| Next 40 TB | $0.085 | $3,400 |
| Next 100 TB | $0.07 | $7,000 |
| Remaining 850 TB | $0.05 | $42,500 |
| Total | $53,800 |
Spread across a month, 1 PB averages only about 3 Gbps. Even allowing for evening peaks, that traffic fits on a pair of 10 Gbps ports, which are billed at a flat monthly rate however many bytes cross them. Committed-use discounts and CDN contracts will lower the cloud figure, so model your own rates, but the shape of the comparison rarely changes.9
Archives grow forever
Mezzanine files, camera originals and multi-language deliverables accumulate into petabytes that are rarely read but must stay accessible. Cold tiers look cheap until you need to restore a season for a remaster or a new territory. On-prem object storage and LTO tape remain hard to beat for deep archive, with a cloud copy kept only for what's genuinely hot.
Live production is latency-sensitive
SMPTE ST 2110, NDI and SRT workflows care about jitter and determinism. Cloud production works well for pop-up events and remote contribution, but a permanent master control room often runs better, and more cheaply, on dedicated infrastructure close to the facility.
| Workload | Profile | Natural home |
|---|---|---|
| 24/7 linear playout | Steady, always-on | Owned or co-located, cloud for DR |
| Live event spikes | Short, massive bursts | Public cloud, burst on demand |
| VOD origin & packaging | Steady, egress-heavy | Colo with direct CDN peering |
| Transcoding backlog | Large, deferrable | Owned baseline, cloud spot overflow |
| Deep archive | Huge, rarely read | On-prem object or tape |
| Audience analytics | Variable, managed services | Public cloud |
AI: the GPU question
AI changes the arithmetic more than anything before it. A training cluster that runs at high utilisation for months is precisely the workload where owning or leasing dedicated capacity can pay back fastest. Meanwhile, inference serving millions of requests a day is steady enough to plan for, and latency-sensitive enough to benefit from running close to users or content.
For media companies, the two worlds now overlap. Automated captioning and dubbing, highlight clipping, metadata tagging, content moderation and personalisation all run AI models directly against the video library. If the library sits on-prem and the models run in the cloud, you pay to move the data every time. Put the compute where the data lives.
Rent or own? A three-year view
Take one 8-GPU H100 server running around the clock. On AWS, the equivalent p5.48xlarge instance lists at about $55 an hour on demand after the June 2025 price cut, or roughly $3.10 per GPU-hour on a three-year Savings Plan.8, 11 Buying a comparable server costs somewhere between $250,000 and $320,000.10
Assumptions: on demand $55.04 an hour; Savings Plan $3.10 per GPU-hour ($24.80 an hour); server bought for $300,000 up front, plus $4,000 a month for colocation, power, support and a share of staff time; 730 hours a month. Excludes networking, storage, finance costs and residual hardware value.
Show the data as a table
| Month | On demand | Savings Plan | Own |
|---|
Run flat out, the owned server overtakes on-demand cloud in month 9 and the three-year Savings Plan in month 22. Over three years it comes out around $1 million cheaper than on demand and $200,000 cheaper than the Savings Plan. Utilisation is the hinge. If the same work only needs the GPUs about 30% of the time and you rent purely on demand, the two options cost about the same. Below that, renting wins.
What to weigh
- Utilisation. If your GPUs are busy most of the day, most days, dedicated capacity is worth modelling. If they sit idle between experiments, keep renting.
- Power and cooling. Modern accelerators need dense power and often liquid cooling. Many existing broadcast facilities weren't built for it; specialist colo and GPU-cloud providers can bridge the gap.
- Supply and refresh. Hardware generations move fast. Plan for a refresh cycle, and consider leasing to avoid stranded assets.
- Data governance. Rights-managed content, talent likeness and training data provenance are easier to control on infrastructure you own.
A note on neoclouds. ‘Exit’ doesn't have to mean ‘own a data centre’. GPU-specialist clouds, sovereign regional providers and bare-metal hosts offer a middle path: dedicated performance and simpler pricing, without building facilities yourself.
Three principles
Portable by default
Containers, Kubernetes, Terraform and open formats. Avoid proprietary services where an open equivalent exists, and wrap the ones you keep.
Data first
Your data is the hardest thing to move. Decide where the master copy lives, and design everything else to come to it.
Hybrid on purpose
Own the steady baseline and burst into the cloud for peaks. Make the boundary a deliberate architectural choice, not an accident.
An illustrative scenario
A mid-sized UK streamer, all in the cloud
Picture a service running four linear channels, a VOD library of 5 PB and a growing set of AI jobs for captioning, dubbing and highlights. Everything sits in one hyperscaler region, and viewers pull about 1 PB a month.
Applying the playbook below, the archive and VOD origin move to a colocation site with direct CDN peering, removing most of the per-gigabyte egress bill. Playout moves next, with the cloud kept as disaster recovery. The AI jobs move to a small owned GPU cluster beside the archive, so the library is never copied out to be processed. Live event spikes, analytics and experiments stay in the cloud.
Figures come from the worked examples above, not from a specific client.
A practical playbook
Whether you're planning a full repatriation or simply want leverage at your next renewal, the sequence is the same.
-
Map every workload
Classify by utilisation, data volume, egress, latency needs and dependency on managed services. Most estates split cleanly once you look.
-
Model true total cost
Compare three years of cloud spend, including egress and support, against hardware, colo, power, network, people and refresh. Be honest about both sides.
-
Break the lock-in
Replace proprietary queues, databases and media services with portable equivalents. Codify infrastructure so it can be rebuilt anywhere.
-
Move data before compute
Use provider exit-waiver programmes for bulk transfer. Establish the new master copy, then migrate the workloads that depend on it.
-
Start with the obvious wins
Archive, steady-state playout, VOD origin and long-running GPU jobs. Keep live-event burst, analytics and experimentation in the cloud.
-
Keep the door open both ways
Retain cloud accounts, DR runbooks and tested burst paths. The goal isn't to leave the cloud for good; it's to make sure you are always able to.
The bottom line
For broadcasters, streamers and AI teams, the public cloud is still a powerful tool. It just isn't the answer to every workload. The organisations that win the next five years will be the ones that can place each workload where it runs best, and move it when the economics change.
A cloud exit strategy is how you earn that freedom. Write it before you need it.
About the author
Sources
- Latham & Watkins: EU Data Act, significant new switching requirements for data processing services (August 2025)
- Competition and Markets Authority: Cloud services market investigation
- Compare the Cloud: CMA opens strategic market status investigation into Microsoft (March 2026)
- CNBC: Google vows not to charge data transfer fees when cloud customers exit (January 2024)
- AWS News Blog: Free data transfer out to internet when moving out of AWS (March 2024)
- CMA: Appendix N, egress fees and free switching programmes (PDF)
- The Register: 37signals expects cloud exit savings of more than $10 million over five years (October 2024)
- AWS News Blog: Up to 45% price reduction for Amazon EC2 NVIDIA GPU instances (June 2025)
- AWS: Amazon EC2 on-demand pricing, data transfer section
- IntuitionLabs: Data centre GPU prices
- Opslyft: AWS GPU pricing
Prices are US list prices as published at the time of writing and change often. Check current rates for your region before making decisions. Nothing here is legal advice.