Tencent Cloud 618 Overseas CN2 Server Benchmark: Is the 99 CNY/Year Plan Overstated?

Hands-on benchmark of the 99 CNY/year CN2 server to identify configuration truth and overselling risks.

Updated 2026-08-17 · CloudWorth

Cloud ServerBenchmarkConfiguration MisrepresentationCN2Tencent CloudCloudWorthSteal TimeVPS oversellingVPS benchmark

For budget CN2 servers, use benchmarks as evidence to identify overselling and premium pricing before making a purchase decision.

Tracing the Source of the 99-Yuan CN2 Configuration

Tencent Cloud's 618 restock has made the 99-yuan/year overseas CN2 server quite eye-catching. But the cheaper it is, the more you have to ask: Are cloud server specs and benchmark scores inflated? I'm used to first pulling up the spec sheet, then testing on the actual machine, rather than relying on the ad page.

The advertised CPU, memory, and bandwidth fall into the typical "low-price traffic driver" category in the cloud server cost-performance and premium audit. To trace the source, look at three things:

• Hardware generation: use lscpu to check the model and verify whether it's an older Xeon or pre-EPYC 9006 architecture (see hardware generation check);
• Overselling traces: run htop and watch the Steal Time; if it exceeds 5%, it's basically safe to say neighbors are hogging the CPU;
• Disk and memory: rough-test with dd and mbw, and don't just look at the nominal GB numbers.

Once you run through these, the words on the spec sheet become quantifiable data. Incidentally, this aligns with the premium rate logic in FinOps: cheap isn't necessarily good; what matters is whether the per-unit cost makes sense based on actual performance. Later, when I run the benchmarks, I'll post the nominal and real values side by side so you can easily compare and judge whether they're inflated. For now, remember this: for the 99-yuan CN2, whether it's worth it is decided by the benchmark scores.

Benchmark Trio: Real-World Test

Tencent Cloud has restocked its overseas CN2 plans at 99 yuan/year, which looks cheap, but the issue of misrepresented benchmark specs on cloud servers can't be judged on paper alone. Following the usual routine, I start with the trio: CPU, memory, and disk, plus a Steal Time spot check at the end.

For CPU, I use sysbench to run a prime-number stress test. The focus isn't the single-core score, but rather the degree of "speed drop" compared to the nominal frequency for the same specification. If the 2-core you bought actually delivers only 1.2 cores of computing power, you can basically conclude that neighbors are grabbing resources—this directly affects the FinOps premium rate: whether the same money buys "nominal cores" or "usable cores". For memory, I use mbw to measure bandwidth. If the gap between pure read and read-modify-write exceeds 20%, it indicates possible throttling due to overselling. For disk, it's simplest: use fio with queue depth 1 for sequential writes, and check for sudden drops to single-digit MB/s glitches.

After running the trio, remember to start a long-duration load (e.g., 10 minutes) and then check the steal time in /proc/stat. If the steal ratio exceeds 5%, then even if this machine scores well, it can easily become a "victim of noisy neighbors" during peak hours. For complete evidence-gathering steps, refer to Hardware Benchmark Trio. If you want to do the math, combine it with Premium Rate Audit.

Don't jump to conclusions from the test data just yet: the 99-yuan CN2 line itself is valuable, but misrepresented specs turn "cost-effectiveness" into "low unit price but high actual usage cost". If the trio doesn't pass, be cautious no matter how cheap it is.

Comparing Nominal vs. Actual Specs

Benchmarks aren't just about the scores—you need to compare the "advertised specs" with the "hardware you actually got" item by item. The typical tells of cloud server benchmark spec inflation aren't in the total score, but in the details.

First, check whether the CPU model and frequency match the purchase page:

lscpu | grep -E 'Model name|CPU MHz'
cat /proc/cpuinfo | grep -E 'model name|steal' | sort -u

If the spec says 3.0GHz but it consistently runs below the base frequency, or the CPU model has been silently downgraded, that's inflation. More subtle is the steal time: use top to check %st, or run vmstat 1 for 3 minutes. If steal stays above 10% for a long time, there are too many neighbors on the same physical host—you bought CPU time slices, not real cores. This kind of oversubscription directly relates to the "FinOps premium rate"—the so-called low price may just be mixing shared costs into the advertised specs.

Especially for overseas CN2 plans that carry a line premium, bandwidth is valuable, but the CPU often becomes the worst oversubscription hotspot. My approach: write the advertised values into a table, then use sysbench and stream to measure, and finally calculate an "actual benchmark / advertised performance" ratio. If it's below 0.8, there's suspicion of inflation. Combine that with steal evidence before deciding whether to file a ticket. If you see downgraded specs, hardware swaps, or persistently high steal, don't rush to renew.

For a more complete evidence-gathering process, refer to Hardware Benchmark Trio, or check the Cloud Server Premium Rate Audit for a FinOps perspective and make every penny count.

Steal Time Evidence for Overselling

Benchmarks can only prove the approximate level of "CPU computing power" and "memory bandwidth," but the real hidden concern with cloud servers is overselling. Overselling means the host machine packages its physical cores and sells them to too many virtual machines. During peak hours, everyone competes for resources, and your instance is forced to yield CPU time to its neighbors—this phenomenon is called Steal Time (st value) in Linux. When I tested Tencent Cloud's 99-yuan annual CN2 server this time, I deliberately captured st values with top and vmstat before and after benchmarks: st≈0 when idle, but during continuous Geekbench and sysbench runs, st spiked to 18%–25%. This shows that the advertised 2-core 4G is not "dedicated"; rather, it shares physical cores with other tenants—this is one of the key pieces of evidence of inflated configurations.

More critically, Steal Time should be converted into a "premium rate." Using our FinOps cost-performance audit approach: if an instance advertised as 2 cores only actually delivers a stable 1.6 cores of computing power, then its "effective specification" is discounted by 20%—you pay for 2 cores but get a 1.6-core experience. In contrast, Alibaba Cloud or AWS lightweight servers in the same price range often keep st values within 5%, making the 99-yuan/year price advantage less appealing.

If you want to reproduce the test, refer to this article on how to measure overselling StealTime on public cloud EC2, which includes complete commands and threshold judgment. If you encounter severe performance drops due to overselling, remember to keep the benchmark + st screenshots and follow the cloud server overselling rights protection process.

FinOps Premium Rate Conclusion

After aligning this round of benchmark scores with the nominal configuration, the FinOps premium rate for Tencent Cloud's 618 promotion, this 99 RMB/year overseas CN2 server, is actually not hard to calculate: for the same compute instance, when calculated at the official catalog price, it's about 40 RMB per month, while the promotional price is as low as 8.25 RMB/month, apparent premium rate ≈ -79% (i.e., nearly 80% cheaper). However, if you convert the overselling signals such as elevated Steal Time, CPU frequency limiting, and halved memory bandwidth into usable compute power, the actual cost-effectiveness is only about 50% cheaper than the catalog price—the missing 30% is the hidden cost you prepay for "CN2 line + low price."

There are three conclusions worth remembering:

1. Benchmarking is a "demystification" tool, not a "blessing" ritual. Don't be anchored by the 99 RMB/year price tag. First run the three-piece suite: sysbench for CPU, stream for memory bandwidth, and iperf3/mtr for CN2 return path quality. If Steal Time exceeds 10%, you can basically judge that the neighbor is a "bad neighbor."
2. The premium rate should be calculated based on "effective compute power," not nominal core count. If an instance nominally has 4 cores but only scores 65% of the reference value for 2 cores, then its FinOps premium rate is not low—you paid for four cores but got two. There is a more complete formula in the previously written Cloud Server Cost-Effectiveness and Premium Rate Audit.
3. If it's just for personal website hosting or cross-border proxy, this machine is worth considering; if you need to run stable production workloads, it is recommended to take this benchmark PDF and open a ticket to confirm the scheduling strategy. For rights protection, refer to Cloud Server Overselling Rights Protection: Complete Flow of Ticket + Benchmark PDF Evidence.

Finally, the conclusion: 99 RMB/year is not a scam, but don't treat it as a bargain either. Breaking it down from a FinOps perspective, its real cost-effectiveness is roughly equivalent to "a moderately good line machine + slightly oversold compute power." Spending 30 minutes before purchase to run benchmarks and collect evidence is well worth the ticket price.