The capacity is real. The 50 MB/s is not

Why buy a card that looks too cheap to be real?

Memory card prices have roughly doubled since early 2025, with NAND flash being pulled toward AI datacenter demand, so a 128GB card that used to be pocket change now is not. Against that backdrop this one sits at roughly half the going rate for its class. That is the single most common question I get about storage listings: is this cheap because it is a bargain, or cheap because it is a lie? The bench answers that question the same way every time, by writing to every last block and reading it all back. It joins CKBench, the running index of every card on the bench, alongside the Botslab 128GB, the FitcamX 64GB and the Kioxia Exceria G2 128GB, which is the card I keep as the corpus benchmark for what a good mainstream 128GB part looks like.

One unit, purchased retail on July 30, 2026. Everything below comes from a single full suite run on that unit, and three passes of a test are an indicator, never a lifetime.

What the box claims, and what the card did

Two rows there are worth pausing on. The card prints no Class 10 mark anywhere, not on the package, not on the card face, and not in the back panel's own model string. That is unusual enough that I checked twice. It does not matter in substance, because the U3 mark it does print asserts the same 30 MB/s floor, but a verdict cannot be given for a claim that was never made, so that row reads NOT CLAIMED even though the card's internal register happily declares Class 10 anyway.

The warranty is the one that would catch a buyer out. The front of the package carries a 1-YEAR WARRANTY badge. The back declines to state a term at all, saying only to check the TEAMGROUP website. Both of the seller's own product listings advertise a lifetime limited warranty covering parts and labor. I re-read the live listing on the day I published this and the gap is still there. The package is the less generous of the two sources, and the package is the thing that arrives in your hand.

Does a cheap card mean a fake card?

Not here, and this is the result that justified buying it. A fake card is one whose controller has been told to report a capacity the flash cannot back. Ask it for 128GB and it says yes, accepts everything you send, and quietly throws away or overwrites whatever will not fit. You find out when you go looking for the files.

The only way to settle it is to fill the entire card with data you can recognize afterward. Every block written here carries a unique seeded stamp, so a controller cannot cheat by noticing repeated data and storing it once. The run wrote 116 files totalling 123,724,890,112 bytes, then read all of it back and compared: zero mismatched bytes and zero read errors. The space is there.

So why does a 128GB card show up as 119GB?

This is the single most common false alarm in memory cards, and it has nothing to do with fraud. Card makers count a gigabyte as 1,000,000,000 bytes. Windows counts it as 1,073,741,824. Same card, same bytes, two different rulers, and the Windows ruler always reads about 7% smaller. A completely honest 128GB card shows roughly 119GB in File Explorer. If yours does, that is not a problem.

What is worth noticing is a second, separate gap. Before any of that arithmetic, this card only exposes 123.72 GB of its 128GB label, which is 96.7%. Every card holds some flash back as spare area for wear levelling and bad blocks, but the amount varies, and this one holds back more than most. On this bench the SanDisk, Samsung, Lexar and ProGrade cards all deliver essentially 100% of their label. The Botslab and VIOFO cards give 97.4%. At 96.7% this is the lowest ratio of any honest card I have measured, about 4.27 GB you paid for and do not get. More spare area is not automatically a fault, and it can even be a reliability choice on cheaper flash. It is simply never mentioned: unlike Kioxia, which publishes an explicit usable figure, this package makes no arithmetic promise at all.

Why does the write speed swing from 5 to 83 MB/s?

Five minutes of continuous committed writing, fan off, sampled once a second. The average is a flat and unremarkable 40.20 MB/s. The individual seconds are anything but: they run from 5.24 MB/s to 82.84 MB/s, which is the widest swing of any card on this bench, two and a half times wider than the Botslab that previously held that title.

It is not instability and it is not a stall. Underneath it is a metronome. Fitting a two level cycle to the 300 samples returns a period of 2.70 seconds: about 1.25 seconds writing at roughly 80 MB/s, then about 1.45 seconds crawling at roughly 6 MB/s, moving about 95 MiB in each fast phase. The fit tracks the real trace closely, and refitting it independently across the 52 minute whole-card write returns the same period every time, holding to within 0.7% from the first gigabyte to the last.

What you are watching is a very small write cache filling and emptying. The card absorbs a burst into fast single-bit flash, then has to fold that data down into slower, denser storage, and while it folds, incoming writes crawl. Most cards do this once, at a cache cliff, after several gigabytes. This one does it 22 times a minute. That is why the cache analysis reports no cliff anywhere in a full-card fill: there is no cliff to fall off, because the card is already at the bottom of it from about the second second onward.

It also explains why every write figure on this card agrees. Sequential 39.76, burst 39.85, steady state 40.20, whole-card fill 40.68. A card with a real cache posts a high burst and a lower sustained number. This one has no burst advantage to give you at all, on anything bigger than about 96 MiB.

Does it still hold the V30 rating it prints?

Yes, and that is the number that decides what the card is safe to record. The verdict here is never an average, it is the worst ten second window of the whole five minutes, because a video recorder does not care what your card averaged, it cares whether it ever dropped below the bitrate for long enough to lose a frame. That window came in at 35.5 MB/s against the 30 MB/s floor V30 requires, so the rating holds with about 18% to spare.

There is a caveat specific to a card that cycles this fast. A ten second window spans only about 3.7 of this card's cycles, so exactly where the window lands nudges the answer: measured at nine, ten and eleven second windows the worst case reads 36.6, 35.5 and 39.5 MB/s. Read 35.5 as "about 36" rather than a figure good to two decimals. The V30 verdict itself is not in doubt at any of those widths, and the standard ten second window happens to be the harshest of them, so the published number errs in the right direction.

Where I would not lean on it is the high bitrate row. The format table clears 4K120 and similar formats by 1.6%, which is inside run to run noise. Treat that as a fail rather than a pass.

Who actually made this card?

Every SD card carries a small identity register called the CID, burned in at manufacture, holding a manufacturer ID, an OEM ID, a product name, a revision and a manufacture date. I read this one on a Raspberry Pi with the card in a native SD slot, so nothing is filtered through a USB bridge. It reports manufacturer ID 0xFE.

That number was already on file here. It is the same manufacturer ID reported by the two no-name cards that ship free in the box with a FitcamX and a Botslab dash cam, and it appears on no public manufacturer list I have been able to reach. A retail card, in a printed hang tag package, sold under a named Taiwanese company's branding, is reporting the same maker ID as the giveaway cards.

On its own that would be weak evidence. An uncatalogued ID could easily be a default value that several unrelated module houses never bothered to change. What makes it stick is a different register entirely. Alongside the CID, every card publishes an SD Status register: not an identity field but a specification sheet the card writes about itself, listing its speed class, UHS grade, video class, app class, allocation unit size and erase timings. Across these three cards, in two different capacities, sold by three different companies, that register is byte for byte identical. Every other card I have read differs somewhere in that block, as parts with different controllers should. These three do not differ anywhere. One field is especially telling: all three declare a PERFORMANCE_MOVE of 4 MB/s where every other card on the bench reports zero.

The write curve agrees, independently, which is what makes it worth stating rather than speculating about. These three cards are the only ones on this bench whose sustained write runs on a repeating two level cycle at all. Fit the same model to the Kioxia or the SanDisk High Endurance and it finds a high level and a low level within 5% of each other, which is the model's way of reporting that there is no cycle there. Two completely different measurements, a register field and a physical behavior, pick out the same three cards. There is more on all of this in the running CID register.

None of that is an accusation. All three cards passed full capacity verification. The finding is narrower and more useful than fraud: the brand on the front of a memory card tells you very little about who built it, and the registers tell you a great deal.

The date the package could not tell me

The same register carries a manufacture date, and this one reads February 2024. The seller's own product page lists the card as first available in April 2026. So this is not new production built to hit a low price during a shortage. It is flash that was made about two years before it was listed, which is a far simpler explanation for the price than anything clever about sourcing. Nothing on the package could have told you that: there is no lot code and no date code printed anywhere on it. Worth keeping in proportion, though, since a manufacture date is a programmed field rather than a measurement, and on a card whose maker ID appears on no list there is no external authority to check it against.

Does it get hot, and does it slow down when it does?

No, and no. With the fan off for the full five minutes it peaked at 133.2 °F (56.2 °C) against a 76 °F (24.4 °C) room, a rise of about 57 °F (31.7 °C). It did not throttle: the write rate at the end of the five minutes matches the start, and the 52 minute whole-card write that followed held the same cycle from beginning to end. In absolute terms that is the coolest 128GB microSD card I have measured.

The rubric asks a harder question than the absolute number, though, which is how much heat a card produces for the work it does. At about 1.6 °F of temperature rise for every MB/s it sustains, this is the worst ratio of any genuine 128GB card here, against 1.1 for the Botslab, 1.1 for the Kioxia and 0.9 for the SanDisk High Endurance. It runs cool because it is barely working, not because it is efficient.

The other half is the card's own rating. 56.2 °C is comfortable against the 85 °C most cards in this corpus are rated to, but this card is rated to 70 °C, which leaves under 14 °C of headroom measured in a mild room. That, rather than speed, is the real reason I would keep it off a windshield.

What does loop recording actually do to it?

Loop recording, and app use generally, is not one long tidy write. It is scattered small writes: file table updates, event saves, overwrites of the oldest clip. The bench approximates that with random 4K writes, which means writing 4 kilobyte chunks to random places rather than one continuous stream, run as three back to back passes so you can see whether the card degrades as its internal map gets churned up.

It holds up well. Across the three passes it measured 481, 434 and 472 write IOPS, which is operations per second: a dip in the middle pass and then a recovery, ending 1.8% below where it started. For comparison, the unbranded Class 4 reference card this bench keeps as its floor collapses by 71% across the same three passes. This card does not do that. In absolute terms 462 IOPS averaged is modest, and it is the reason the A1 row in the claim table cannot be settled here at all: the A-class specification is measured with command queuing that USB card readers do not perform, so an A1 result on this bench would be measuring the reader, not the card.

Worth being clear about what this does not prove. Three passes over a minute is a churn indicator. It is not an endurance test, and nothing in a single session can tell you whether a card will still be working in three years. This card also prints no endurance rating of any kind, so there is not even a manufacturer claim to check it against.

How the Newegg TeamGroup 128GB scores

The bottom line on the Newegg TeamGroup 128GB

It is real, it is cheap, and it is slow. Those three things are the entire review. I bought this card expecting to find something wrong with it, because a 128GB card at roughly half the going rate during a flash shortage is exactly the profile of a scam. Nothing is wrong with it. Every byte of the space is there and verified, it holds every speed class mark it prints, it never stalled once across a full-card write, and it does not throttle when it gets warm.

What you give up is real too. It is the slowest genuine 128GB card this bench has measured, it misses its own printed write speed by 20%, it hands back less of its label than any other honest card here, and it carries no endurance rating and a one year warranty against a lifetime one advertised online. For a phone, a tablet, a handheld console, a drone or a 4K60 camera, none of that will ever bother you and the card is a perfectly sensible buy. For a dash cam, a security camera, or anything that needs the number on the front of the box, buy the card that actually posts it.

If you want a straight upgrade in the same capacity, the Kioxia Exceria G2 128GB sustains 62.6 MB/s at its worst against this card's 35.5, and the SanDisk High Endurance 128GB sustains 88.0 and is rated to 85 °C. Both cost more. Both are more card.

Disclosure: I bought this card at retail with my own money. Newegg and TEAMGROUP had no involvement in this test and no editorial input, and neither knew it was happening. This post contains affiliate links. As an Amazon Associate I earn from qualifying purchases, at no extra cost to you. Affiliate links like these and my YouTube channel are what keep the site running.