By clicking “Accept”, you agree to the storing of cookies on your device to enhance site navigation, analyze site usage, and assist in our marketing efforts. View our Privacy Policy for more information, or manage your preferences.
We've been building one since 2020. Ten million sessions later, here's the math nobody runs before they start — and the long tail of work nobody warns you about until you've already hired the team.
No marketing tricks. No fake totals. Numbers below are sourced and dated; assumptions are disclosed.
Should you build or buy a virtual classroom platform in 2026? Almost always: buy.
Building a production-grade virtual classroom — one that runs at 99.99% reliability across every browser, every device, and every school network, with the compliance posture a district will sign off on — costs ~$30M over three years with a 50-person engineering team, across 80 distinct production components. The hard parts aren't the WebRTC or the SFU; they're SOC 2 Type II evidence (which has to age in time), native iOS and Android apps, the testing matrix across browsers and devices, and 24×7 on-call. Even Class Technologies, after raising $160M+ in venture funding, built their UX on top of Zoom rather than build their own real-time video.
Pencil Spaces, by comparison, handles this from $3,000/month on Scale (and from $6/month on our entry tier) — for the same SFU, whiteboard, recording, transcription, AI summaries, native mobile, scheduling, identity, and compliance posture. We've been operating it in production since 2020, with 10M+ sessions delivered across customers in 60+ countries at 99.95% measured uptime.
Read the full math: the $30M anchor · the 80-component stack · the 50-person team · infrastructure economics · the CPaaS / BBB shortcut math · the interactive calculator.
To stand up a virtual classroom platform that's roughly comparable to what you can buy from us today — running at 99.99% reliability across every browser, every device, every school network. Three years. Fifty people you'll need to hire and retain. Infrastructure that bills by the second. Compliance that takes longer to pass than your cap table takes to close.
We get this email a couple of times a month. Almost always from a smart engineer who just shipped a polished prototype over a weekend — Cursor in one window, a managed video SDK, a transcription API, a CRDT library, an auth provider. It works. They feel ~80% done.
We've been that engineer. So have a lot of our customers, before they became our customers. So this isn't a lecture — it's the thing we wish someone had said to us in 2020.
If you've read that and still feel certain your team is the exception, run the calculator below with your numbers. We'll happily lose the argument if your model produces a smaller number than ours — and if it does, we'd genuinely like to hear about it. We learn from the engineers who pull this off; they're rare, but they exist, and they're not the buyers we're talking to here.
The first — and largest — cost overrun on every internal video build is the gap between the Friday-afternoon whiteboard estimate and the actual surface area of a production system.
We'll grab an open-source video SDK, drop in a whiteboard, plug it into our LMS, and ship it in a quarter. Two engineers and a designer. Maybe $300K?
Not wrong about the components. Wrong about everything else — the percentage of work each one represents, the operational tax of running them, and the second-derivative cost of maintaining them once you have real users.
You can ship a working prototype in a quarter. You cannot ship a platform a tutoring program will sign a multi-year contract for, in less than three years.
The naive list is roughly 15% of the work. The other 85% is where startups die — quietly, six quarters in, after the founders have already told the board it's “almost done.”
These are the actual systems we run in production. Skip any one of them and you ship a prototype, not a platform. Most teams discover them after their first outage.
What it actually takes to run a system at 99.99% reliability across every browser, every device, every school network. The asterisks aren't on whether you'll hire these roles — they're on whether you'll be able to retain them in a market where every one of these humans is being recruited weekly by every other company that's serious about real-time.
Public list pricing from each vendor as of 2026. Assumes a modest scale of 10,000 user-hours per month after Year 1 — the rough size of a 200-tutor program running five hours a day. Bigger operations get worse ratios.
Edtech buyers will not sign a contract without these. School districts have a procurement checklist; tutoring programs are inheriting it. Each line below has a real audit, a real lead time, and a real invoice.
Every one of these is a specific incident we, or peers running production WebRTC, have personally debugged. They are not in the architecture diagram. They are why real-time video has been called the hardest commodity-grade software to ship.
Audio device selection nondeterministically picked the wrong input. Reproduced in production for ~3% of mobile users. Took two senior engineers six weeks to root-cause — the bug was in WebKit, not our code. Workaround required redesigning our audio capture flow.
You don't fix this with a Stack Overflow answer. You fix it by building enough logging infrastructure to see one bad call out of every thirty.
Hardware AEC quality varies by chipset. Some devices ship aggressive noise gates that cut quiet voices entirely. Some Xiaomi MIUI builds bypass standard audio APIs. Mid-range Samsung devices in India have tested differently than Samsung devices in Korea.
You ship a software AEC fallback. You build a device-specific config table. You buy twenty cheap Android phones and keep them on a shelf. You hire someone whose job is this table.
H.264 vs. VP8 vs. VP9 vs. AV1 negotiation across browsers, OS versions, and hardware acceleration paths is a maze of partial support. Older Safari only does H.264 baseline. Some Firefox builds require an AV1 fallback. The spec says one thing; implementations do another.
You write a compatibility matrix. You maintain it. You re-test it every Chrome release. The matrix lives in your codebase and silently grows.
Tutoring is a 4–7pm business. Every weekday, your concurrent session count jumps 8× in 20 minutes. You discover that your SFU autoscale group warms slower than the surge, that your Postgres write queue has a hot row, that your recording cluster runs out of file descriptors.
None of this is visible in dev. It is visible at 4:08pm Eastern when a tutor can't get into a session and emails the founder.
Real student. Real session. The connection survives because of bandwidth estimation, adaptive bitrate, retransmission, FEC, and a custom jitter buffer that we tuned for six months on a Brazilian cellular profile we'd never seen before.
On a vibe-coded MVP, this session drops at the 90-second mark. The student gives up. The tutoring program loses the contract. You never know.
Many K-12 networks restrict UDP entirely or rate-limit aggressively, which kills standard WebRTC. You need TCP fallback through TURN-over-TLS on port 443 — the same port as HTTPS — or your sessions don't connect at all.
This isn't optional. It's an entire customer segment. You discover it the first time you do a district pilot.
AI cannot answer this email. AI cannot SSH into the affected pod, read the logs, notice that a transcoder worker has wedged on a malformed audio packet, restart it, and send a reassuring reply at 11:23pm.
That requires a human. That human is on your payroll. You did not put them on your spreadsheet because you didn't know yet that this email exists.
Chrome ships a major release every six weeks. Most are quiet. Some break your code. You don't see this in your CI; you see it on a Wednesday morning when your error rate goes from 0.4% to 9%.
You read every Chrome release note. You subscribe to webrtc-discuss. You file bugs upstream. This is permanent overhead and it never reduces.
Adjust the sliders to your situation. The model uses Bay Area senior comp, public list pricing for infrastructure, and 2026 vendor rates. Try lowering everything as far as the model allows — the number doesn't get small.
We use AI every day. We ship faster because of it. And we are the team most qualified to tell you what AI cannot do for a production virtual classroom — not because it's incapable, but because the work it can't do is the work that matters.
AI takes the prototype from weeks to weekends. It does not take the platform from years to months.
Every dollar and engineer-month you sink into rebuilding undifferentiated infrastructure is a dollar and engineer-month not going to your actual product. Same starting line. Two different finishes.
Where most teams who chose to build are, on month thirty-six.
Where teams who bought infrastructure and built differentiation are, on month thirty-six.
These aren't projections. They're the operating envelope of the platform you can buy from us today — running in production, every day, since 2020.
I've been the founder on the other side of this conversation. In 2020 my co-founder Amogh and I walked away from senior tech-leadership roles at companies like Meta and Google to make the bet on building a real-time, virtual-classroom-grade platform from first principles. The decision wasn't wrong — we had a thesis about what tutoring needed that nothing on the market did. The cost — in years, in headcount, in the long tail you read above — was real.
The thing nobody told us — and the thing I now tell every founder who asks — is that the infrastructure isn't the product. It's the part of the product you're forced to ship before you get to ship the part you actually care about.
Six years later, that infrastructure is what we sell. As a full-stack platform, called Pencil Spaces. As an embeddable API, called Carbon. We built it once, at unsustainable cost, because we had to. You don't.
If you want to build the differentiated product on top of it — the curriculum, the matching engine, the reporting layer, the AI workflow — we'd love to hear about it. That's where the actual moat is. The video plumbing isn't.
We're not the only company that's tried to build a virtual classroom platform from scratch. Two of the most prominent venture-backed attempts of the last five years, and what they had to raise to ship something:
Read that second card carefully. Class raised more than five times this page's build-cost anchor — and didn't even build their own real-time video. They wrapped Zoom and built UX on top. Even that was a $160M+ undertaking. Engageli, building closer to the metal, raised $47.5M and runs leaner. Neither story is a knock on the founders — both companies have credible operators. It's a knock on the assumption that this category is cheap to ship. It isn't. The receipts above are receipts; the receipts on this card are the receipts the rest of the industry has filed.
Sources: public funding announcements, Crunchbase, Tracxn, PRNewswire (Class Series B, July 2021). Verify before relying.
Two paths, depending on whether you want a finished product or the infrastructure underneath one. We sell both, because we built both.
The full virtual classroom, branded for your program, ready in a week. Everything in the stack diagram — we run it; you teach.
The same infrastructure, exposed as an API. Embed real-time video, whiteboard, and persistence into your own app. Skip the build.
Two paths come up on every Scale call. Each looks dramatically cheaper than building from scratch — and is, by ~5%. Here's where the other 95% reappears.
The pitch is reasonable: instead of operating your own real-time media stack, you pay a managed vendor by the participant-minute and they handle the SFU, the TURN, the scaling, the codec wars, the surge load on the third Tuesday in October. Public list pricing across the category, as of 2026:
At our reference scale of 10,000 user-hours per month — that's 600,000 participant-minutes — the CPaaS line item lands around $2,000–$2,400/month at the category-standard $0.004/min, or ~$1,000/month on Chime SDK. Year one: ~$25K. Cheap. Cheaper than running your own SFU cluster, by a meaningful margin.
Now what that doesn't include. Recording is separate. Transcription is separate. The whiteboard is separate. Mobile SDKs are separate. Compliance is separate. Scheduling, billing, identity, support tooling, reporting, integrations — all the other 79 components in §02 — are separate. The CPaaS replaces cluster 01 of sixteen. You still owe the other fifteen, and you still owe the engineering team in §03 to build them, integrate them, and operate them.
Pencil Spaces handles this workload from $1,200/month on our entry tier ($6/month base + $0.12 per user-hour overage) — less than the CPaaS would charge for just the video minutes. Scale tier, the right fit for production programs, starts at $3,000/month (custom-quoted at higher volumes) and includes the SFU plus the other 79 components: whiteboard, recording, transcription, AI summaries, mobile apps, scheduling, identity, compliance posture, on-call, and the integration glue between all of it. The arithmetic is plain: same money or less than the CPaaS for ~80× the surface area, and the strategic-direction risk is ours, not yours.
BigBlueButton is a credible, mature open-source virtual classroom. We use it in tests. We respect the project. It's a good answer to a specific kind of question: “can we run small-scale online classes on a budget without paying a per-seat fee?” Yes. You can. We're not going to pretend otherwise.
It's a worse answer to the question this page is actually about, which is: “can we operate a serious tutoring program at 99.99% reliability across every browser, every device, every school network, with the compliance posture a district can sign off on, without funding a build?” The honest cost stack at moderate scale (200 concurrent users, modest customization, K-12 use case):
BBB is “free” in the same way a sailboat is free if a friend hands you the keys: the boat is real, the wind is real, and so is everything you didn't budget for — the slip, the insurance, the haul-out, the surveyor, the diesel, the time. Most teams that try this path underestimate the DevOps line and the compliance line. Both are load-bearing. Neither is negotiable.
Pencil Spaces Scale starts at $36K/year for this workload (custom-quoted at higher volumes), full-stack, modern UX, mobile apps shipped and maintained, compliance handled, no fork to merge against, and a roadmap that moves at your pace because the platform is purpose-built. Our entry tier — Pro at $6/month + $0.12 per user-hour overage — runs smaller programs for less than $15K/year, full-stack. The “free license” framing is one of those headlines that doesn't survive a CFO's spreadsheet.
We're not pretending these shortcuts don't exist or that they never work. CPaaS works for teams whose differentiation is exclusively in the application UI, who are happy to treat video as commodity infrastructure, and who can absorb vendor-strategy risk. BBB works for teams with a DevOps function to spare, whose UX needs map closely to BBB defaults. For everything else — especially K-12 programs that need to clear SOC 2, FERPA, native mobile, and district procurement — the math doesn't favor either path. We've watched it not favor them, repeatedly, for six years.
It usually isn't. Here are the most common reasons buyers tell us they think the math above doesn't apply to them.
We'll handle the video, the whiteboard, the compliance, the on-call.
You build the thing nobody else can.
Questions about your specific situation?hello@pencilspaces.com
