There are three ways to arrive at this estimate:  looking at John Blankenbaker's words, considering his parts purchasing process, and looking at serial numbers.  However, none of these methods gives a clear number.

John's own words are imprecise as he seems purposely vague.  The best quote is his 2007 oral history, when he talks about selling out to CTI and shipping remaining parts and computers to South Carolina.  He said "I think I probably delivered a few computers to them because I had sold about 44 out of that 50 and so I delivered some to them" (emphasis added.)  Why is he being so vague?  This is uncharacteristic for a meticulous engineer.  He's likely imprecise because of the discrepancy of how many he sold compared to how many he made (which would include his prototype,  the serial number 183 which was not working until 2010, the serial number 216 that he kept until 1986 and donated to the computer museum, or the one he "gifted" (or gave as partial pay) to his employee of 4 months.

If we take his statement at face value, that he actually sold 44 computers, and shipped the remaining to CTI, we get an unreasonable number.  We know that CTI had at least 8 computers (because Robert Nielsen purchased all 8 of his computers directly from CTI.)  Nielsen was certain CTI never made any from scratch, and thought CTI never sold any machines other than to him.  This would make 44+8 = 52 total computers, then we add the three computers he kept for himself (the prototype, Serial #216 which he donated to Boston Computer Museum, Serial #183 he fixed in 2010) and the machine he gifted to his employee of 4 months,  we count a total of 56 computers.  This number is unlikely since it's high above his "run of 50" he ordered parts for.  It makes much more sense if the 44 number he completed, which would include the three he kept, and the one he gifted, and maybe a few complete that he sent to CTI?  This is hard to reconcile.

A better way to estimate the number of Kenbak-1s made is to consider his parts ordering process.  When John designed and built the first prototype board, he likely purchased parts in small quantities --- likely enough for only one or two computers.  He probably wanted some spares in case a part was defective or damaged.   Only after he got the prototype working, he contacted investors, and raised about $20,000, using about $12,500 to order parts for 50 machines (each machine cost about $250 in parts.)  This was because many parts had discounts in quantities of 50.  Therefore, it makes sense that he ordered enough parts for 51 or 52 computers all together.  Perhaps this explains where his approximation of 44 came from:  maybe 52 total sets of parts, 8 sent to CTI, so perhaps he made (rather than sold) about 44 including his prototype.  This does make a little more sense.

It may be best to consider the most unique and and most costly computer part:  the printed circuit board.  He says he made only a single copy of the prototype PC board in his oral history ("I could have built them on the basis of that first board. It wasn’t that bad but it would have been very desirable to change it. I’d only bought one to start with so—".)  After he got the prototype ("Rev ___") working, he apparently ordered at least two copies of the fixed board ("Rev A".)   After he found that it had a few remaining mistakes he again modified the PC board design to the "Rev B" boards.   The question then is:  Did he order a just a few samples of the "Rev B" boards to check if they were right,  then have to order an additional 50 to get a 50-unit price?  John's comments in the Felsenstein interview suggests it was a slow time for the PC board company, and the costs of boards wasn't very much (about $30 each) so I suspect they would allow him some flexibility with quantity.  It's possible he got as few as 48 of the "Rev B" circuit boards, or as many as 50 (if he had to get a full 50.)  This would make a total of 51 or 53 PC boards made (one prototype, at least two "Rev A's", and 48 or 50 "Rev B's") all together.  Lastly, we know he kept one bare PC board which never had any components soldered to it.  It was photographed many times after 2005, loaned to Grant Stockley for his Series 2 reproduction, and even loaned to Achim Baque to copy.  So with one unused PC board, it seems logical that 50-52 PC boards were available to make into computers.

Finally, people often count serial numbers to determine manufactured quantities.  Find the highest and lowest serial number,  subtract them, and add one.  But I don't think this works for the Kenbak-1.  He started his first serial number at #167 (from his address at 12167 Leven Lane in Los Angeles) and the highest known serial number was #216 (from the computer he donated to the Boston Computer Museum in 1986.)  This all too conveniently suggests a series of exactly 50 within that interval, so if we add the "non-serial numbered prototype", we have a total of 51.  But it seems several computers were not serial numbered.  And the 50th serial number (#216) on the museum donated computer is a bit suspicious.  While this 50th serial number would suggest it was the last one made, the internal "Rev A" board actually proves it was one of the very first couple made, and the small area of the PC board where the serial number was usually stamped has an odd blemish I haven't seen on any other PC board, almost suggesting a prior serial number mark was wiped off.  Is it possible John just placed the 50th number sticker (#216) on the computer he donated for symbolism?  Regardless, if we took the serial number evidence at face value, we again have 50 computers made, plus the prototype which never had a serial number, making a total of 51.

There are several reasons for variation.  First, different instructions take different amounts of time to execute.  Most instructions are two bytes, and some instructions with indirect addressing modes require many steps and memory accesses to execute.  Since memory is implemented serially, when reading a byte from memory, there may be a shorter or longer delay until the required byte circulates to the output of the 1024-bit shift register to be read.  In fact, the vast majority of execution time on a Kenbak-1 is spent waiting for a required memory byte to be read from the serial memory.  Lastly, each Kenbak-1 has a different clock frequency.  Quartz oscillators were not used.  Simple RC circuits , with imprecise capacitors were used, and those components change as they age.

The best way to know, execution time is empiric measurement of a typical program on a real machine.  When running the simple "counting" program (on the first page of the kenbak.com web site) on an original computer (Nielsen #3 or "Grey Case") the high-order bit flashes on and off almost exactly once per second.  This corresponds to 768  instructions per second.  Unfortunately, after 50+ years this Kenbak-1 doesn't have a clock speed of 1 MHz -  it is now around 1.6 MHz.  This may be due to aging of capacitors or even the carbon composition resistors.  If the clock speed was decreased to 1 MHz (by adjusting resistors and capacitors on the multivibrator clock) it follows that the instruction rate would be around 480 instructions per second.

A second Kenbak-1 can also be similarly measured.  A YouTube video shows John Blankenbaker running a similar counting program on the Prototype Kenbak-1 in 2018.  (https://www.youtube.com/watch?v=JNU6PFJj4tg)   John entered the counting program by memory, so it's probably the same counting program except the "A" register was incremented rather than "B" register.  This program is published in his "Laboratory Exercises" manual, as the very first program a student enters and runs.  The YouTube video shows the high-order bit turning on and off 7 times in about 11 seconds.  That would correspond to about 488 instructions per second.  That suggests it's clock rate is quite close to the  nominal 1 MHz.

One more computer was then measured:  a "Kenbak-1 Series 2" reproduction computer made in 2007 (see kenbakkit.com)I.  While this is a modern reproduction, it uses the same circuitry and components as the original computers, and it seems to be operating at 428 instructions per second.  I have not yet measured it's clock rate, but suspect it's around 0.9 MHz based on ratios of speed..

Inspecting some of the circuits on the earliest Kenbak-1 computers seem to show some variation in capacitors used for the multivibrator clock circuit.  The first prototype computer seemed to use two different capacitors for the otherwised symetrical clock circuit, a big brown mica capacitor, and a large ceramic disc capacitor.  The Nielsen3 Revision A computer seems to use two big brown mica capacitors.  The next made John2 computer in the Computer History Museum uses two large ceramic disc capacitors.  But then all other production Revision B computers seem to show very small ceramic disc capacitors.  It does seem John was playing around with different capacitors on his first few computers, before settling on a certain value, and this may add to variability between machines.  

After years of studying the Kenbak-1, I have opinions, and they differ a bit from those of John Blankenbaker.

I think schools were resistant to purchase the Kenbak-1 because machine code programming is not as good for introductory programming as higher-level languages.  John thought the problem was the slow and complex school budgeting process, and perhaps that's the excuse the schools told him.  John did a good job of marketing to schools.  He traveled to educator conferences to demonstrate his computer.  He put advertisements in popular education magazines and journals.  John correctly recognized that timesharing systems were his main competition in schools.  Timesharing systems allowed schools to lease terminals and phone lines to connect to a large central computer serving many schools and businesses.  He acknowledged this competition in an advertisement in November 1971  Nations Schools which compared the cost of buying a Kenbak-1 to the higher cost of renting time sharing terminals.  But I suspect the issue wasn't the cost, it was the capability.  Timesharing systems offered high level languages, such as Dartmouth BASIC, or COBOL which are much better for teaching high school students than machine code.  A student can make a much more useful programs with BASIC than they could with machine code, using real numbers, not just 8-bit bytes.  Machine or assembly code programming is usually taught later in computer science or engineering courses, and it is harder to solve real life problems.  So while John recognized his competition, I don't think he realized why schools were choosing his competition.  

Why the individual customer (known as the hobby, or amateur customer) didn't buy is more complex.  I suspect the level of detail in advertisements or sales brochures was probably too limited for individuals to shell out what was a huge amount of money (equivalent to $5,500 in 2022.)  John thought he didn't focus on this hobby market enough, but he actually did reasonably well.  He got an early article written in the Amateur Computer Society newsletter and his advertisements in "Scientific American" and "Computer World" were undoubtably seen by thousands of interested people.  He said he even put on a demonstration of the Kenbak-1 for the Homebrew Computer Club, where Steve Jobs and Steve Wozniak were members, long before the Apple-1 was invented (but John wasn't sure if Steve and Steve attended that day.)  But it would be quite an act of faith (or naivety) for an individual to send a check for $750 by mail after seeing only very limited description in these advertisements.  While marketing through magazine advertisements was common in the 1970's, many mail-order products weren't half as good as their advertisement suggested.  Many "toy computers" were commonly advertised, including the "Geniac" or "Braniac" in the mid-1950's or even the "Digi-Comp 1" or "Digi-Comp 2".   One example of a "toy computer" is this item at the "oldcomputermuseum"  which was sold at the same time for $30 (about $250 in 2022 dollars) which was just $5 worth of parts and cardboard, and probably disappointed hundreds of buyers.  John's advertisements focused on how "fun and educational" the Kenbak-1 was, but lacked enough details for customers feel confident they knew what to expect.  It's not at all surprising that only about 6 individuals bought his computers, and that only included about 3 engineers or programmers (numbers from the Amateur Computer Society newsletter.)

Had John sold 200 or 400 kits the first year, history might have been much different.  He could have hired more help to assemble the computers and handle administrative details.  Then John could have done what he did best: focused on designing accessories and improvements.  Maybe even make a second generation "Kenbak-2" using the microprocessor and high-density RAM chips which were coming out soon.

For the 1986 Boston Computer Museum contest the industry expert judges realized their first job was to decide what qualifies as a "Personal Computer."  Even defining what a "computer" is was not obvious.  They had to decide if they should include programmable calculators, analog computers, and even the cardboard or plastic "toy computers" like the "Geniac" which were commonly sold from the 1950's through the early 70's.  The judges decided not to include programable calculators, and the mechanical, or analog computers, focusing only on digital computers with a stored program.  They even made the arbitrary decision to disqualify any "kit" computers, and focus on items which were commercially available.  Many people have offered opinions of what a "personal computer" should be.  The Deutsches Museum in Munich Germany saw the ambiguity in definition, so placed a note on their Kenbak-1 saying it is the "first personal computer for educational purposes."  But that does little to clarify the ambiguity. 

In November 1999, an early and thoughtful treatise tried to name the "First Personal Computer."  The "Blinkinglights.org" site (permanently archived here) asked "what was the first personal computer?"  They do an excellent job of explaining the difficulty in defining a PC, and listed several contenders, usually ruling them out.  Unfortunately, they go down the rabbit hole of including the cardboard and wire "toy computers" like the Geniac, the paper-clip computer, and others, before finally settling on the 1950 "Simon" which most people would not consider was a true computer, and definitely not a stored program computer that people think of today.  Instead of convincing readers that "Simon" was the first personal computer, many readers were left feeling the Kenbak-1 matched their own notion of a first personal computer.    

While only around 51 were made, almost all Kenbak-1s were sold to high schools, colleges, and technical schools.  So many hundreds of students from 1971 to 1976 and beyond used these to get their first exposure to computer programming.  This was a pivotal time when every high school wanted to start their own computer classes.  But in addition to the hundreds who learned to program on a Kenbak-1, thousands of other people saw the ubiquitous advertisements or heard about this first real computer for under $1000 and were inspired to imagine a world of affordable computers.  The Kenbak-1 left a legacy of interest and imagination which eclipsed it's very dismal production numbers.  Even 50+ years later, many enthusiasts and hobbyists continue to put together kits and reproductions and learn to program a Kenbak-1, trying to get a deeper understanding of one of the ancestors of our modern computer.