(originally posted on LinkedIn February 26, 2024)
Introduction
Now that funding has begun, the CHIPS Act is back in the
news, with the White House and the Department of Commerce’s CHIPS Program Office
(CPO) announcing $5B allocated for R&D projects and $1.5B to
GlobalFoundries. Coincidentally, just a few days before these announcements, I
was asked my opinions of a the proposals in a report titled “Reshoring
and Restoring: CHIPS Implementation for a Competitive Semiconductor Industry”. This report was published by an organization
called the American Economic Liberties Project (AELP). I hadn’t heard of the AELP before, but
according to one of its recent job listings, it’s a “non-profit 501(c)3
organization dedicated to fighting monopoly power”.
In a nutshell, through its publications and other
activities, AELP means to foster aggressive antitrust policies through the use
of legislation, litigation, federal regulation and oversight, tariffs, and
other means, with a special focus on the technology industry. While I support reasonable
efforts to ensure competitive markets and counteract economic inequality, after
reading the report, I hesitated to bring attention to this group and its work, owing
to the numerous erroneous assumptions and factual errors I found in it. Even worse, the authors present a defamatory characterization
of leading semiconductor companies and their business practices, as well as the
entire fabless-semiconductor industry.
My first inclination was to not give such a report any further
exposure. But someone who closely monitors activities in Washington brought it
to my attention, and my subsequent research showed that AELP is very well
connected there, having been founded by a former Treasury official now working
with the FTC. According to the website Influence Watch, the group’s 2023
Anti-Monopoly Summit was attended by a number of federal officials, and the
agenda included a video message from President Biden. Along with that, the
report’s lead author is an academic and former mobile-industry executive now running
for political office. Unsurprisingly, there’s not a single
semiconductor-technology expert among the report’s authors; the other two are a
lawyer and a political scientist.
After realizing that this report might make its way into the
hands of people in power, I decided to publish a more extensive critique here,
where semiconductor-industry colleagues can add their expertise and comments on
AELP’s proposals. This work is entirely unsponsored.
But first, I contacted the report’s lead author, inviting him
to discuss what I see as its mischaracterization of the semiconductor industry,
based on my 45+ years of experience in roles throughout the ecosystem. I didn’t
think it would be fair to blindside him with criticisms that came out of the
blue. As you can read below, that didn’t
go well.
I expect that the AELP isn’t the only such organization
seeking to use the CHIPS Act to further its agenda, and while its proposals
aren’t quite Sam Altman scale, the group is at least partially funded by a
billionaire—Pierre
Omidyar—founder of eBay. The report and its proposals serve as a great
example of what can happen when idealogues, politicians, and lawyers attempt to
make technology policy. I believe
readers will see that the AELP’s policy proposals are the product of its
confirmation bias, rather than sound economic and technological reasoning. My
rebuttal got much longer than I expected, because the more I read the 61-page
document, the more falsehoods I discovered ... but you can skip all that and just
scroll down to the end to judge the proposals for yourself.
The AELP Takes on TSMC
The report starts with what the author’s call the “Monopoly
Of Leading-Edge Logic Chips”, based on a description of TSMC’s dominance of
the foundry business. It’s well known that concern for reliance on chip
fabrication in Taiwan is the primary motivation for the CHIPS Act. To explain
how this happened, looking at TSMC’s history and the evolution of the fabless
industry, one could identify a number of reasons why the Taiwanese company
leapt into the forefront in chip fabrication.
One might point to more than thirty years of innovation, and
the development of a vast collaborative chip-design ecosystem that includes
advances made by EDA and IP vendors, along with the breakthroughs in wafer-fabrication techniques created
by equipment suppliers. Looking at the tremendous growth of that ecosystem, one
might also conclude that pure-play foundries have been proven to be more efficient
than those owned by IDMs, such as Intel and Samsung, which have fallen behind
TSMC.
But the AELP authors see things differently. According to
their report, these are the reasons for TSMC’s position on the leading edge of
semiconductor fabrication:
“Two decades of patent abuse,
exclusive deals, cheap money, and weak antitrust enforcement have gutted what
used to be a crown jewel of the American economy. While there is a need for
significant capital investment to sustain the semiconductor business, this
consolidation is the result of business practices, not economies of scale.”
That last point really stood out… the notion that the
success of the fabless-semiconductor industry, and TSMC’s leadership, have
nothing to do with economies of scale. And innovation fell? I have no idea how
they determined that!
“If economies of scale were
driving the needs of this industry, semiconductor firms would be building fabs
to compete — as they used to do and as certain parts of the industry still do —
instead of paying out shareholders.”
“If Apple, Qualcomm, or Broadcom
operated in a competitive market, then margin pressures would compel them to
vertically integrate…”
The AEPL apparently endorses that famous quote from Jerry
Sanders, “real men have fabs”. They’re just now $20B ones.
In their attacks on the fabless model, the authors attempt
to use as a comparison how memory chips are manufactured, concluding that the
IDM model employed by companies like Micron Technology is more efficient, and
that “Where the semiconductor industry is highly competitive, the fabless
model is incompatible.” So, according to AELP, IDM = competitive, fabless =
anti-competitive. DRAMs, flash memory, and logic chips all require advanced
process technology, but the authors failed to comprehend how different those
design and manufacturing requirements are, as well as how different the end
markets are. Here’s a more appropriate comparison: in 2023, Micron reduced its
CAPEX plan by 40%, to $7B, whereas TSMC’s first 3nm fab was estimated to cost
$20B and its Phoenix fab will cost an estimated $40B.
Making Apple the Culprit
The AELP authors also blame Apple for making TSMC so strong,
going so far as to say that without them, it would be unlikely that the foundry
could “retain its key fabs in the hot zone of Taiwan”. Then, reading
into the headlines about Apple acquiring all of TSMC’s 3nm production in 2023,
AELP concludes that:
“The Apple-TSMC deal excludes
competitors from acquiring advanced semiconductors and from building
competitive fabs elsewhere”
This, after noting that Intel cancelled its 3nm orders, so
they weren’t actually excluded. Nevertheless, TSMC only began booking 3nm
revenue in 3Q23, when it contributed just 6% of the foundry’s quarterly income.
At the same time, AMD, Qualcomm, and Nvidia (as well as Apple) were employing TSMC’s
4nm technology, and are expected to roll out 3nm chips in 2024. Also in 3Q23,
Mediatek announced it had taped out its first 3nm design, for 2024 production
at TSMC. But to the report’s authors:
“It would be as if one plane
manufacturer bought the entire supply of aluminum for a year.”
Not even close. Apple is just the lead customer for TSMC’s
3nm node. Rather than a case of excluding competitors, Apple accepted the risks
inherent in early production, in return for small increases in performance (10-15%
speed) and density (35%) compared with 4nm, and at best a few-quarter lead over
competitors. Keeping in mind that a full mask set for TSMC’s 3nm process is
estimated to cost $20M, it should be apparent it’s not something very many
companies can afford.
But to AELP, Apple is a monopoly, by virtue of its 60% share
of the U.S. smartphone market. That doesn’t meet the definition of a monopolist,
and it’s an even weaker argument when taking into account that Apple’s worldwide
smartphone-market share is only around 30%. Nevertheless, the report’s authors errantly
assert that:
“Apple’s monopolization of
smartphones means that TSMC has effectively one customer for the world’s most
important chip category.”
TSMC has just one smartphone-chip customer? The authors omit
the fact that Qualcomm (holding about a 28% share of the smartphone-chip market)
has historically split production of its Snapdragon processors between Samsung
and TSMC, and Snapdragon processors ship in many more phones than Apple’s
iPhones, as do TSMC-manufactured Mediatek chips.
AELP vs The Fabless-Semiconductor Industry
As I mentioned in the introduction, before deciding to
publish this review, I first contacted the report’s lead author. By offering to
discuss the report, I thought I could share my knowledge to help correct some
of its errors, and address the misunderstandings of the semiconductor industry.
Suffice it to say, that didn’t go well. As an example, twice I asked the author
how he justified one of the report’s most defamatory allegations, which he
ignored:
“Fabless designers compete based
on the strength of patents and exclusivity of their commercial agreements,
instead of the institutional capacity of their productive facilities, their
ability to deliver predictably to customers, and the excellence of their engineers.”
Besides being a blatantly false characterization of competitive
dynamics in the fabless-semiconductor industry, I take this statement as an
insult to all semiconductor engineers, past and present, and the countless
innovations they have created. In my brief exchange of messages with the
author, I pointed out other examples of erroneous assumptions and factual
errors, but got no response to those either.
Instead, I was told that I missed the point, and “There
is no question that the semi industry is more concentrated and less competitive,”
and that’s that. He claimed that rather than looking at the entire
semiconductor industry (despite devoting more than 20 pages to it), the report
is meant to address this:
“A handful of firms with market
power over specific semi sub-segments stifles innovation and is a massive tax
on everything downstream.”
But that’s not at all the way the report reads. There are
numerous statements in it that paint the entire industry with an accusatory
brush, going well beyond describing a “handful of firms”. Here’s
another:
“The U.S. semiconductor
industry used to be competitive and vibrant, but today it is too
concentrated and is plagued by many of the dangers one would expect from a
consolidated sector, notably shortages, weak innovation, high prices,
and the pooling of risk by a monopolist.”
Once again, weak innovation?
The industry “used to be competitive and vibrant”? In recent
years, through my work as an analyst, I’ve spoken with at least fifty startup
companies. That claim of weak innovation and a lack of vibrance is just
nonsense. But to the AELP, the growth of the fabless-semiconductor industry is
the result of lax regulation, the failure of policymakers, and financial
pressures from Wall Street, which has enabled chip suppliers to create
monopolies and “to avoid the expense of physical production”. This ignores the fact that the fabless
ecosystem has enabled an unprecedented wave of innovation, and the creation of hundreds
of new companies that wouldn’t exist if they had to build their own fabs.
Playing a Game of Concentration
To support its claim that the U.S. semiconductor industry
has become too concentrated, the AELP authors claim:
“Since 2010, intra-sector
acquisitions have shrunk the number of independent U.S. semiconductor firms by
over 40%. Concentration and profits rose while innovation and resilience fell.”
Here’s how the AELP authors came up with their measure of
industry concentration: by removing from an odd list of 27 semiconductor
companies (the “Top U.S. Chip Makers” in
2010), the 12 that have since been
acquired or merged (12/27 = 44%). That’s very clever math, but not at all a
measure of industry concentration. There’s no comparison whatsoever of the
number of U.S. semiconductor companies in 2010 versus 2024.
Looking at some of the companies on AELP’s list of so-called
Top U.S. Chip Makers, one is PMC-Sierra, which had multiple rounds of
layoffs and had restructured just prior to its 2016 acquisition by Microsemi. Another
company on the 2010 list, Atmel, was struggling as the 7th-place MCU
supplier before Microchip bought it in 2016. The list also includes industry
pioneer Fairchild, after recording a loss of $172M in 2016 – it was acquired by
ON Semiconductor. The weak don’t survive in any market.
To actually measure industry concentration, it would be appropriate
to compare the share of industry revenue of the top companies then and now,
rather than the totally false and misleading picture that forms a key pillar of
AELP’s argument. According to industry statistics, in 2010, the top 10 U.S.
semiconductor companies accounted for 55.5% of the U.S. industry’s revenue,
whereas in 2023 that number was 40.8%.
But to the authors of this report, it’s as if every
logic-chip segment leader holds a monopoly:
“Fabless firms monopolize
specific categories of logic chips: Qualcomm for wireless modems; Nvidia for
GPUs; AMD and Intel for CPUs; and Broadcom for network and broadband chips”
Qualcomm has a monopoly in modems, so Mediatek, Samsung,
Huawei... throw out your designs. AMD and Intel are co-monopolists in CPUs,
more accurately a duopoly then, and they don’t compete? Forget about all the
Arm-based processors in non-PC segments of the computing market, or the
progress Arm is making in servers, such as Microsoft’s announcement of its Cobalt
chip. Broadcom is notorious for its run-ins with the FTC, but that hasn’t
exactly put Cisco and Marvell out of business. In fact, Marvell’s 2023 revenue
growth rate quadrupled Broadcom’s, and now Broadcom is up against Nvidia’s surging data center networking business.
While it’s well known that Nvidia is dominant in this early
stage of data center AI, it faces growing competition in that arena from a
number of companies, and the AI-market overall—especially inference—includes
many more competitors. Any claim that Nvidia’s lead is by virtue of anything
other than its innovations is absurd. But since it has a monopoly in GPUs,
throw away your AMD Radeon cards.
[The report also includes a ridiculous discussion of how
Nvidia GPUs will cause Intel “server CPU mass extinction”, but I’ll
leave that for anyone motivated to read it. It’s clear that the authors don’t
understand how CPUs are still required in conjunction with GPUs, to handle AI
as well as general-purpose workloads.]
Some of the alleged monopolists’ competitors do get a brief
mention later in the report, in a table of logic-chip market leaders. But
that’s well after the authors establish their monopoly argument, and whatever
competition they acknowledge for leading-edge chips is reduced to being called
weak and highly concentrated.
“Despite there being many logic
chip firms, any given logic market niche is highly concentrated and often
dominated by a single firm”
The cost of working in leading-edge technologies demands
targeting large markets, not niches, but that’s just poor use of the
terminology. This interpretation of
concentration is fundamentally flawed. As mentioned above, a 3nm mask set costs
$20M, and the cost of each wafer is estimated to exceed $20,000. Only a few companies that possess superior
talents and budgets can afford the most-advanced process technologies. Of
course there are relatively few companies at the leading edge, just as there
are relatively few players at the leading edge of any market, whether it’s
semiconductors, high-performance sports cars, fine wine, luxury goods… you name
it. Despite what the AELP says, this is entirely natural. It’s called market
segmentation. There will always be more
companies serving mainstream and commodity segments, and the fewest at the top
of the pyramid producing premier products.
AELP’s Proposals for CHIPS Act Implementation
Having built its case for how concentrated and lacking in competition
and innovation the semiconductor industry is, here are the remedies that the
AELP proposes, along with my comments. Some of the proposals directly relate to
technology, others to economic policy, and the rest to government oversight of
private companies.
1. Incubate
entrants with a goal of four independent, leading-edge logic foundries
This may look like a sensible idea in theory, until you look
at the practical challenges of developing leading-edge semiconductor-manufacturing
technologies. The AEPL authors propose
prioritizing funding for “second-tier foundries”, so they can “upgrade
their capacity and sophistication over time”.
Since TSMC is already building a fab in Arizona, the AELP proposes
that Apple should fund that. Whereas the authors describe TSMC as a monopoly
for leading-edge technology, I assume that second-tier means every other foundry
in the U.S. That leaves Intel, Samsung, and GlobalFoundries, but GF decided to
stop developing leading-edge processes at the 7nm node. Upgrading “over time”
would take many years. At its recent foundry event, Intel announced a goal of
being “the world's No. 2 foundry”... by 2030.
Regardless of the real-world obstacles, the author’s concept
of “independent” but “leading-edge” fabs is essentially a
contradiction in terms, because they expect these foundries would be
interchangeable. As any chip designer knows, no two processes at two
independent foundries are exactly the same, even if they use the same node name
(e.g. 3nm). Each has its own PDK (process design kit), a unique process recipe,
with different design rules, different gate density, performance, power
efficiency, yield, and so on. And at the cost of a leading-edge fab, building
four of them would exceed the funds available under the CHIPS Act.
In the coming years, I don’t have any doubts that we’ll have
more advanced chip fabs in the U.S. But leading edge is just that… one company
will always be the leader in some respect, there can’t be four exactly the same. Perhaps Intel and Samsung will catch up with
TSMC, offering more options for manufacturing leading-edge designs. But this
only gets harder and more expensive with each succeeding process node.
2. Pass
legislation to require all fabless firms to dual-source their manufacturing
“Ideally the largest fabless chip companies (Apple,
Qualcomm, AMD, Nvidia, etc.) would be required to buy from multiple foundries and
commit to purchase at least 30% of their global volume from U.S.-based fabs.”
The AELP favors government control
such as this. But just like the proposal above, it shows their lack of
knowledge and naivete regarding semiconductor technology. Setting aside the question
of how the government can or should monitor a semiconductor company’s volume
and manufacturing contracts, second sourcing at the leading edge means more
than doubling the design cost. It would mean two different tapeouts, two
different mask sets, twice the design-verification effort, double the chip qual
effort. Because some of the chips I
designed went into military applications, the DoD would require a second
source, but that was for chips with basically one function, like an
analog-digital converter. Even then, the two chips were never identical. For
leading-edge SoCs, there’s just no way. As in the first proposal, leading-edge
foundries aren’t interchangeable.
3. Include
Federal Trade Commission to allocate CHIPS funding and set criteria for funding
recipients
This is just adding more government
bureaucracy, based on the AELP’s belief that “The semiconductor market is
already concentrated and thin, with limited buyers and sellers in any given
segment.” Mergers and acquisitions already go through FTC review. But AELP
wants the CPO and FTC to also take into account ”any future plans for
intra-sector acquisitions” of CHIPS-funding recipients.
Besides being unrealistic, this
proposal is completely irrelevant to the CHIPS Act. Funding under that program only
goes to manufacturers, not the fabless chip-design companies that the AELP
accuses of antitrust activities.
4.
Develop thicker markets
a)
by protecting new fabs through long-term
demand contracts from both government and private-sector buyers
b)
developing a set
of chip-making standards that lower switching costs and promote
interoperability
c)
ensuring foundries operate under a
quasi-common carrier principle
On the first point, the AELP is
proposing the government get involved in the terms of private-companies’
purchase contracts. The third is just as
absurd, based on the author’s mistaken belief that Apple locked out competitors
from using TSMC’s latest 3nm process. It proposes that the government intervene
in how a foundry allocates it supply.
But point ’b’ once again
shows how politicians and lawyers are completely unqualified to make technology
policy. The notion that a government
agency could define a “set of chip-making standards” for any fabrication
process, not to mention a leading-edge one, is ridiculous. Any semiconductor
professional knows that.
5.
Address the long-term incentives for the underlying
fabless business model by:
a)
by requiring more open patent and IP
licensing practices and
b)
disincentivizing extractive financial
practices, like buybacks and dividends
Point ‘a’ is based on the
author’s erroneous allegation that “Fabless designers compete based on the
strength of patents…. (instead of) their ability to deliver predictably to
customers, and the excellence of their engineers.” The authors propose that “CHIPS funding
recipients should be held to stringent requirements to openly and fairly
license their patents”. But again, CHIPS funding only goes to foundries,
not chip-design companies, so this proposal is irrelevant.
6.
Increase most-favored nation tariff rates
on end-use chip products
With this proposal, the authors
intend to address offshore assembly of finished products, mainly consumer
products. The authors believe that increasing import tariffs will stimulate the
creation of a more diversified OSAT (outsourced semiconductor assembly and
testing) industry, causing “some portion” to move to the U.S.’s
free-trade partners in the Western hemisphere. Perhaps, but only by raising
costs to consumers.
7.
Legislation to create an American
Semiconductor Supply Chain Resiliency Fee (ASSCR).
This proposal targets the supply chain for mature
process nodes, to prevent the “possibility that Chinese or other foreign
producers may dump their excess chip capacity on American markets.”
“We propose that this could be
done through a tax on mature-node chip products that rely almost exclusively on
foreign-sourced chips”
“This new fee — the American
Semiconductor Supply Chain Resiliency (ASSCR) Fee —
would require original equipment manufacturers to either source a certain
percentage — for example, 30% — of their products’ semiconductor value added
from U.S.-based fabs or pay a marginal but non-negligible fee on all products
with a retail price of $300 or higher, on the order of $10-$20 dollars per
device.”
The authors go on to say that “Chip
products could be defined as any final product where a minimum value-added of
chips is used in its manufacture. The term should be defined to ensure the
inclusion of the key products using mature-node chips: smartphones, cars, and
other consumer electronics.”
So, to implement such a tax, a
government agency would need to review the BoM for every single electronic
product that sells for more than $300. Looking just at modern automobiles now
employing thousands of chips, and EVs increasing the amount of semiconductor
content, this proposal for government review is completely unrealistic.
8.
Pass legislation to establish demand-side
subsidies for electronics manufactured with domestic chips
As an alternative to the ASSCR fee,
the AELP authors propose tax credits to consumers, or other incentives to OEMs
for the purchase and production of products made with U.S.-manufactured chips.
Finally, the authors acknowledge the logistical difficulties of ensuring
domestic content, but they deem it necessary to “break the stranglehold that
a handful of semiconductor firms have over the U.S. economy”. I think such
a hyperbolic statement requires no further comment.
Conclusions
My analysis of the AEPL report began with a fairly typical
request from a semiconductor-industry outsider, for my expert opinions of its
proposals for CHIPS Implementation. As a
semiconductor-industry veteran, what were my takes on its recommendations,
especially its proposed mandate for dual-sourcing, incubating new entrants, and
incorporating FTC oversight? Those
proposals were relatively easy to dismiss, but then I discovered the report’s
defamatory characterization of the semiconductor industry, going so far as to
accuse leading chip suppliers of illegal activity. This comes from an
organization that has tight connections in the federal government. I couldn’t
let that go without further exposure.
The issues that the CHIPS Act are meant to address are too
important for any group of ideologues to exploit. I believe that it should be
obvious that this report is the result of a strong confirmation bias, but
misinterpretations are one thing, deliberately misrepresenting the facts is
another. The author I contacted rejected any criticism or corrections, and that
is especially worrisome.
Fortunately, the good news is that organizations being put
in place to administer CHIPS projects are being staffed with experienced
semiconductor-industry professionals. One
such organization is Natcast, a non-profit entity created to operate the
National Semiconductor Technology Center (NSTC) consortium, established by the
CHIPS Act. Its executives and trustees include individuals with experience at IBM,
Intel, MIT Lincoln Labs, Synopsys, and Zilog. Hopefully, more knowledgeable
insights will prevail over those of biased idealogues, lawyers, and
politicians.
